<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Daniel (San) Ávila</title>
    <description>The latest articles on DEV Community by Daniel (San) Ávila (@dani_avila7).</description>
    <link>https://dev.to/dani_avila7</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F976708%2Fd07d7618-b828-43f9-8031-1e69e7ec0eb7.jpeg</url>
      <title>DEV Community: Daniel (San) Ávila</title>
      <link>https://dev.to/dani_avila7</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/dani_avila7"/>
    <language>en</language>
    <item>
      <title>api-security-audit maps every endpoint in your API before it looks for a single bug</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Thu, 19 Nov 2026 15:53:27 +0000</pubDate>
      <link>https://dev.to/dani_avila7/api-security-audit-maps-every-endpoint-in-your-api-before-it-looks-for-a-single-bug-noi</link>
      <guid>https://dev.to/dani_avila7/api-security-audit-maps-every-endpoint-in-your-api-before-it-looks-for-a-single-bug-noi</guid>
      <description>&lt;h2&gt;
  
  
  Why the inventory comes before the bug hunt
&lt;/h2&gt;

&lt;p&gt;Most security passes over an API are a pattern search. You grep for &lt;code&gt;eval&lt;/code&gt;, for string-built SQL, for &lt;code&gt;jwt.decode&lt;/code&gt;, and you get a list. That list will never contain the two bugs that actually leak an API: broken object level authorization and broken function level authorization. Nothing about &lt;code&gt;findById(req.params.id)&lt;/code&gt; looks wrong on its own. It is wrong only next to the sibling handler that scopes the same query to the caller.&lt;/p&gt;

&lt;p&gt;So the first thing &lt;a href="https://www.aitmpl.com/component/agents/security/api-security-audit" rel="noopener noreferrer"&gt;api-security-audit&lt;/a&gt; does is build an endpoint inventory. One row per method and path, with the handler at file and line, whether authentication applies, which ownership or tenant check runs, which input the caller controls, and whether the operation is sensitive. It follows every &lt;code&gt;use&lt;/code&gt;, &lt;code&gt;include_router&lt;/code&gt; and &lt;code&gt;mount&lt;/code&gt; on the way, including the &lt;code&gt;/v1&lt;/code&gt; router nobody remembers mounting, and notes middleware order, because a guard registered after a router does not protect it.&lt;/p&gt;

&lt;p&gt;Then it reads the rows against each other. A &lt;code&gt;GET /things&lt;/code&gt; that filters by the current user sitting next to a &lt;code&gt;GET /things/:id&lt;/code&gt; that does not is the finding, and that finding only exists as a comparison.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2123268852196786589-695" src="https://platform.twitter.com/embed/Tweet.html?id=2123268852196786589"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2123268852196786589-695');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2123268852196786589&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Install it, then ask for it by name
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest --agent security/api-security-audit
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That writes &lt;code&gt;.claude/agents/api-security-audit.md&lt;/code&gt; in your project. Project agents win over &lt;code&gt;~/.claude/agents/&lt;/code&gt; on a name collision, so an installed copy overrides your personal one.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flcxstk74o2xvj44hrwvv.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flcxstk74o2xvj44hrwvv.png" width="800" height="499"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The frontmatter is four lines and both of the interesting ones are worth knowing:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Read, Grep, Glob, Bash&lt;/span&gt;
&lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;inherit&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;tools&lt;/code&gt; is an allowlist. The agent gets those four and nothing else, no MCP tools unless you name them, which is how a security auditor stays unable to edit your code. &lt;code&gt;model: inherit&lt;/code&gt; means it runs on whatever model the main conversation is on, so an audit does not silently drop to a cheaper model than the session you started.&lt;/p&gt;

&lt;p&gt;Claude Code routes to it from the &lt;code&gt;description&lt;/code&gt; field plus your request, so "audit the security of our API" usually lands. For a guarantee, name it: "use the api-security-audit subagent on &lt;code&gt;src/routes&lt;/code&gt;". Worth knowing that &lt;code&gt;/agents&lt;/code&gt; is no longer a creation wizard; it now prints a reminder to edit &lt;code&gt;.claude/agents/&lt;/code&gt; directly.&lt;/p&gt;

&lt;h2&gt;
  
  
  What comes back
&lt;/h2&gt;

&lt;p&gt;The agent runs in its own context window and the main conversation only gets its final message, so you get a report, not a transcript of the search. The report has a fixed shape:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;One line: how many endpoints mapped, how many findings, by severity.&lt;/li&gt;
&lt;li&gt;Findings, most severe first. Each one carries the OWASP category (&lt;code&gt;API1:2023&lt;/code&gt; and friends), &lt;code&gt;file:line&lt;/code&gt; plus variants, the exact request an attacker would send, the missing check in a sentence or two, and the fix as a specific code change.&lt;/li&gt;
&lt;li&gt;The condensed endpoint inventory, with the vulnerable rows marked.&lt;/li&gt;
&lt;li&gt;“Checked and fine”: the risky-looking places it examined and why they are safe.&lt;/li&gt;
&lt;li&gt;“Worth a look” and “Not checked”.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Items 4 and 5 are the ones I would not have thought to ask for. “Checked and fine” tells you what coverage you actually got, and “Not checked” is where gateways, WAFs and anything outside the repository go, instead of being quietly assumed away.&lt;/p&gt;

&lt;p&gt;Before any of that, every candidate goes through a refutation step: trace the whole path to the sink, look for a global guard or an ORM default scope that already blocks it, confirm the route is really mounted, and name the attacker as unauthenticated, any logged-in user, or a user of another tenant. If an admin is needed to attack an admin feature, it is not a finding. The bar is 80% confidence that it is exploitable; weaker suspicions get one line under “Worth a look”.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it will not do
&lt;/h2&gt;

&lt;p&gt;The “do not report” list is as much of the design as the checks are. It will not tell you to add rate limiting everywhere, add a WAF, or add security headers to a JSON API. It will not report outdated dependencies with no reachable vulnerable call, example secrets in &lt;code&gt;.env.example&lt;/code&gt;, or MD5 used for an ETag. Regex and recursion denial of service is out, with the narrow exception of unbounded page sizes, uploads and paid-API calls per request.&lt;/p&gt;

&lt;p&gt;It is also read-only by construction. Fixes come back as patches in the report, each with the regression test that fails while the hole is open, and the main conversation applies them if you say so.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2122012847958007962-202" src="https://platform.twitter.com/embed/Tweet.html?id=2122012847958007962"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2122012847958007962-202');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2122012847958007962&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;When it does run something, it is bounded: only &lt;code&gt;localhost&lt;/code&gt;, only after checking that &lt;code&gt;DATABASE_URL&lt;/code&gt; and the config point at a local or throwaway database, with outbound side effects such as email, SMS, payments and webhooks stubbed first. No traffic to production or staging, no brute force, no real credentials it finds in the repo. Each finding says whether it was proven by running it or by reading the code.&lt;/p&gt;

&lt;p&gt;One caveat the install does not tell you: listing &lt;code&gt;Bash&lt;/code&gt; in &lt;code&gt;tools&lt;/code&gt; makes the tool available, it does not grant permission. Subagents inherit the parent conversation permission rules, so your &lt;code&gt;settings.json&lt;/code&gt; deny rules still apply to everything this agent runs. If your project denies Bash broadly, the agent falls back to reading code and says so under "Not checked".&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2115255736561234301-950" src="https://platform.twitter.com/embed/Tweet.html?id=2115255736561234301"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2115255736561234301-950');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2115255736561234301&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;A second caveat, on the agent file itself: its &lt;code&gt;description&lt;/code&gt; carries three example blocks, and the combined descriptions of all your subagents have a 15,000 token budget before Claude Code warns you at startup. It is a warning, not a hard cap, but install a dozen agents with descriptions this long and that is where the startup message comes from.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the Agents live
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fl174dwtkq1p5t24glep6.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fl174dwtkq1p5t24glep6.png" width="799" height="468"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The agent: &lt;a href="https://www.aitmpl.com/component/agents/security/api-security-audit" rel="noopener noreferrer"&gt;https://www.aitmpl.com/component/agents/security/api-security-audit&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Its two hand-offs, both in the same category: &lt;a href="https://www.aitmpl.com/component/agents/security/supply-chain-security" rel="noopener noreferrer"&gt;supply-chain-security&lt;/a&gt; for dependency and supply chain risk, and &lt;a href="https://www.aitmpl.com/component/agents/security/compliance-auditor" rel="noopener noreferrer"&gt;compliance-auditor&lt;/a&gt; for SOC 2, PCI DSS and HIPAA mapping. The audit agent names them under “Not checked” instead of pretending to cover them.&lt;/li&gt;
&lt;li&gt;The rest of the catalog: &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;https://www.aitmpl.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>security</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Jev Skill Suggestion keeps your Claude Code skills out of context until one is needed</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Wed, 18 Nov 2026 15:52:50 +0000</pubDate>
      <link>https://dev.to/dani_avila7/jev-skill-suggestion-keeps-your-claude-code-skills-out-of-context-until-one-is-needed-4m86</link>
      <guid>https://dev.to/dani_avila7/jev-skill-suggestion-keeps-your-claude-code-skills-out-of-context-until-one-is-needed-4m86</guid>
      <description>&lt;p&gt;Claude Code sends the model a skill listing on every session: one line per installed skill, name and description. Each entry can run to 1,536 characters, and the whole listing is capped at 1% of the context window. If you have thirty skills installed and your prompt touches one of them, you are still paying for thirty.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2101885477158547753-539" src="https://platform.twitter.com/embed/Tweet.html?id=2101885477158547753"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2101885477158547753-539');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101885477158547753&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.aitmpl.com/component/mod/productivity/jev-skill-suggestion" rel="noopener noreferrer"&gt;jev-skill-suggestion&lt;/a&gt; removes that listing and puts a small decision model in front of every prompt instead. The model picks at most one skill, or none, and only that skill’s SKILL.md reaches the context window.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Jev is
&lt;/h2&gt;

&lt;p&gt;Jev is TypeSafe’s System One decision model. It does not write text. You give it a state and a set of typed questions, and it answers with calibrated probabilities in a few hundred milliseconds. Three primitives cover everything it does:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;noul&lt;/code&gt;: a yes or no question, answered with the probability that the answer is yes&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;choice&lt;/code&gt;: pick one option out of a labeled set, answered with a distribution over all of them&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;score&lt;/code&gt;: a graded value on a scale you define&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The Mod uses &lt;code&gt;choice&lt;/code&gt; to rank your skills and &lt;code&gt;noul&lt;/code&gt; to ask whether the prompt needs a skill at all. You do not write any of that yourself. It ships configured.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2101485309095514241-450" src="https://platform.twitter.com/embed/Tweet.html?id=2101485309095514241"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2101485309095514241-450');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101485309095514241&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Install it
&lt;/h2&gt;

&lt;p&gt;You need Claude Code 2.1.278 or newer and the function hooks flag.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="kd"&gt;export&lt;/span&gt; &lt;span class="nv"&gt;CLAUDE_CODE_ENABLE_FUNCTION_HOOKS&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;1
npx claude-code-templates@latest --mod productivity/jev-skill-suggestion
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That is the whole install. With no API key it falls back to the engine’s own classifier, so it does something useful before you sign up for anything. Add a TypeSafe key or a Vercel AI Gateway key later, when you want the better decision.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fv1v516uk49ggudkzheaa.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fv1v516uk49ggudkzheaa.jpeg" width="800" height="408"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Then run &lt;code&gt;/jev-skill-suggestion:setup&lt;/code&gt;. It reads your installed skills, shows you the plan, asks before it touches anything, and writes a backup you can restore. After that, &lt;code&gt;/context&lt;/code&gt; should count your skills at 0.&lt;/p&gt;

&lt;h2&gt;
  
  
  The settings worth touching
&lt;/h2&gt;

&lt;p&gt;Everything has a working default. These are the ones you might actually change:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;alwaysListed&lt;/code&gt;: comma separated skills that stay visible to the model whatever the prompt. Use it for the one or two you always want in reach.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;neverSuggested&lt;/code&gt;: skills Jev is never offered. They still run when you type &lt;code&gt;/name&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;gateThreshold&lt;/code&gt;: how sure the model has to be that the prompt needs a skill at all. Default 0.3. Raise it if you get suggestions on questions that were only questions.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;timeoutMs&lt;/code&gt;: the latency budget, default 800. Past it, the prompt goes through with no suggestion.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;They go in &lt;code&gt;pluginConfigs&lt;/code&gt; in your settings, or through &lt;code&gt;/config&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"pluginConfigs"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"jev-skill-suggestion"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"alwaysListed"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"commit-helper"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"gateThreshold"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;0.45&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"timeoutMs"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;800&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every failure path is fail-open. If a request errors or runs past the budget, the prompt goes through untouched. It never blocks you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is the trade worth it
&lt;/h2&gt;

&lt;p&gt;I benchmarked it against Claude models doing the picking instead: 26 skills, 10 prompts, three runs each.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Jev: 90% accuracy, 330ms median&lt;/li&gt;
&lt;li&gt;Haiku: 80%, 1.24s&lt;/li&gt;
&lt;li&gt;Sonnet, Opus and Fable: 100%, between 2.2s and 4.5s&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2102167460136796594-439" src="https://platform.twitter.com/embed/Tweet.html?id=2102167460136796594"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2102167460136796594-439');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2102167460136796594&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;So it is not the most accurate option. It is 4 to 14 times faster than asking a model, and it is the only one of the five that also keeps the skill list out of the context window, which was the point. With a handful of skills none of this matters. Past twenty, the listing is real money on every session.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the Mods live
&lt;/h2&gt;

&lt;p&gt;Both Jev Mods are on &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt;, with the install command, the full options table and the source for each one. This article covers &lt;a href="https://www.aitmpl.com/component/mod/productivity/jev-skill-suggestion" rel="noopener noreferrer"&gt;jev-skill-suggestion&lt;/a&gt;. Its companion, &lt;a href="https://www.aitmpl.com/component/mod/productivity/jev-model-router" rel="noopener noreferrer"&gt;jev-model-router&lt;/a&gt;, puts the same decision model in front of the model and effort choice on every turn, and the two run together.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftj01i5t895pu156qng3f.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftj01i5t895pu156qng3f.jpeg" width="800" height="408"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The rest of the catalog, Mods, Agents, Skills, Commands, Hooks, MCPs and Settings, installs the same way from &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Let Jev pick the model and the effort for every Claude Code turn</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Tue, 17 Nov 2026 15:52:56 +0000</pubDate>
      <link>https://dev.to/dani_avila7/let-jev-pick-the-model-and-the-effort-for-every-claude-code-turn-1d5b</link>
      <guid>https://dev.to/dani_avila7/let-jev-pick-the-model-and-the-effort-for-every-claude-code-turn-1d5b</guid>
      <description>&lt;p&gt;Picking a model is a decision you make once, at the start of a session, and then live with for two hours of work that was never uniform to begin with. Renaming a variable and designing a migration do not need the same model, and they certainly do not need the same reasoning effort.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;jev-model-router&lt;/code&gt; moves that decision to where it belongs: per task, made by something that is good at deciding and bad at everything else.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2101176629745561686-319" src="https://platform.twitter.com/embed/Tweet.html?id=2101176629745561686"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2101176629745561686-319');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101176629745561686&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Jev is not a chat model
&lt;/h2&gt;

&lt;p&gt;Jev is TypeSafe’s System One decision model. It does not generate text. You send it a state, which is any string or JSON you have lying around, plus one or more typed questions, and it returns calibrated probabilities for each one in a single round trip. Your code does the thresholding. The model never decides anything, it only measures.&lt;/p&gt;

&lt;p&gt;There are three primitives, and picking the right one is most of the work:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;noul     yes or no      returns the probability the answer is yes
choice   which one      returns the pick plus the full distribution
score    how much       returns a probability weighted value over your levels
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The request is one POST to &lt;code&gt;api.typesafe.ai/v1/systemone&lt;/code&gt; with &lt;code&gt;model&lt;/code&gt;, &lt;code&gt;state&lt;/code&gt; and a map of &lt;code&gt;questions&lt;/code&gt;. Every question in the map is evaluated in parallel, which is why asking three costs about the same as asking one.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2101485309095514241-524" src="https://platform.twitter.com/embed/Tweet.html?id=2101485309095514241"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2101485309095514241-524');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101485309095514241&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  The three questions the router asks
&lt;/h2&gt;

&lt;p&gt;This is the whole classification, straight from &lt;code&gt;hooks/policy.ts&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;questions&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;provider&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Provider&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Record&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kt"&gt;unknown&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;tier&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="kd"&gt;type&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'choice'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;instructions&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'Which is the cheapest tier that can complete this coding task well?'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;criteria&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;TIER_CRITERIA&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;effort&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="kd"&gt;type&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'score'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;instructions&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'How much step-by-step reasoning does this task need?'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;criteria&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;EFFORT_RUBRIC&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;risky&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="kd"&gt;type&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;provider&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'typesafe'&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="s1"&gt;'noul'&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'boolean'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;instructions&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="s1"&gt;'Carrying out this task would itself change production, move real money, or alter data that cannot be restored.'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three primitives, one per question. &lt;code&gt;tier&lt;/code&gt; is a choice across three descriptions of the work: mechanical and local, ordinary engineering, or hard and high stakes. Note what is missing. The decision model never sees a model name. It rates the work, and the mod maps the tier to &lt;code&gt;fastModel&lt;/code&gt;, &lt;code&gt;balancedModel&lt;/code&gt; or &lt;code&gt;deepModel&lt;/code&gt;, which default to haiku, sonnet and opus.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;effort&lt;/code&gt; is a score on a four level rubric. &lt;code&gt;risky&lt;/code&gt; is a noul on TypeSafe's API and a boolean on the Gateway, the same question under two names.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two bars, because the two mistakes are not symmetric
&lt;/h2&gt;

&lt;p&gt;A router that treats “spend more” and “spend less” as the same decision is a router that will eventually hand a migration to haiku. So there are two thresholds:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;minUpgradeConfidence&lt;/code&gt;, default &lt;code&gt;0.3&lt;/code&gt;. Spending more on a task that did not need it costs money.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;minDowngradeConfidence&lt;/code&gt;, default &lt;code&gt;0.6&lt;/code&gt;. Spending less on a task that did need it costs you the task.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;risky&lt;/code&gt; above &lt;code&gt;0.7&lt;/code&gt; goes to the deep tier with real reasoning, past both bars. That one is not a confidence question.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A backend that reports no confidence at all may only move a request up. Spending less on an unmeasured hunch is the bad trade, so the mod refuses to make it. And every failure, a non 2xx, a timeout past &lt;code&gt;timeoutMs&lt;/code&gt; (800 by default), a malformed body, leaves the request exactly as Claude Code built it. The router never blocks a turn.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it hooks in, and the one switch that is off
&lt;/h2&gt;

&lt;p&gt;The mod is a plugin built on function hooks, and it registers three events:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;prompt.submit   classify the prompt before the turn starts
turn.step       apply effort (and model, if enabled) to the turn
agent.spawn     pick the model for each subagent
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three switches control what it is allowed to rewrite:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;routeSubagentModel   model of each subagent      default on
routeMainEffort      effort of the main loop     default on
routeMainModel       model of the main loop      default off
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The last one is off on purpose. A subagent starts with its own context, so routing its model costs nothing beyond the classification. Changing the main loop’s model mid session invalidates the prompt cache, and on a long context the re-cache can cost more than the cheaper tier saves. Turn it on after you have measured your own sessions.&lt;/p&gt;

&lt;p&gt;With &lt;code&gt;logDecisions&lt;/code&gt; on, you see both the raw answer and what the policy did with it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;[jev-model-router] jev: tier fast (0.87) · effort 0.4 → low (0.71) · risky 0.02 · 249ms
[jev-model-router] main loop → effort low: fast (confidence 0.87)
[jev-model-router] jev: tier fast (0.41) · effort 0.4 → low (0.38) · risky 0.01 · 210ms
[jev-model-router] main loop: kept opus/medium, wanted haiku/low (confidence 0.41)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That last pair is the router working correctly and declining to act. It wanted to spend less and did not clear the bar.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest --mod productivity/jev-model-router
&lt;span class="nv"&gt;CLAUDE_CODE_ENABLE_FUNCTION_HOOKS&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;1 claude
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It needs Claude Code 2.1.259 or later. &lt;code&gt;--mod&lt;/code&gt; writes the plugin into &lt;code&gt;.claude/skills/jev-model-router/&lt;/code&gt;, and Claude Code only reads a project's &lt;code&gt;.claude/&lt;/code&gt; once you accept the trust prompt on a first interactive run.&lt;/p&gt;

&lt;p&gt;Then the key, in user settings, not project settings:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"pluginConfigs"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"jev-model-router@skills-dir"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"options"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"typesafeApiKey"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;""&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The entry name is the plugin id, and the id depends on how it loaded: &lt;code&gt;jev-model-router@skills-dir&lt;/code&gt; from &lt;code&gt;.claude/skills/&lt;/code&gt;, plain &lt;code&gt;jev-model-router&lt;/code&gt; with &lt;code&gt;--plugin-dir&lt;/code&gt;. Under the wrong key every option silently stays at its default and the log says &lt;code&gt;no key set&lt;/code&gt;. A Vercel AI Gateway key works too, under &lt;code&gt;gatewayApiKey&lt;/code&gt;, with one caveat: the Gateway's answer carries no confidence field, so confidence is read from an optional distribution and the downgrade bar often cannot fire.&lt;/p&gt;

&lt;p&gt;With no key at all the mod still runs, falling back to Claude Code’s own &lt;code&gt;classify&lt;/code&gt; with the small fast model.&lt;/p&gt;

&lt;p&gt;The order to do this in:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Install the mod and start Claude Code with &lt;code&gt;CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Confirm the &lt;code&gt;ready on&lt;/code&gt; line names the backend you expect. If it says the built in classifier, your key is under the wrong &lt;code&gt;pluginConfigs&lt;/code&gt; id.&lt;/li&gt;
&lt;li&gt;Leave the defaults alone for a few sessions and read the &lt;code&gt;jev:&lt;/code&gt; lines.&lt;/li&gt;
&lt;li&gt;If it keeps declining downgrades you would have taken, lower &lt;code&gt;minDowngradeConfidence&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Only then consider &lt;code&gt;routeMainModel&lt;/code&gt;, and watch your cache stats after you do.&lt;/li&gt;
&lt;li&gt;Run &lt;code&gt;claude plugin validate .claude/skills/jev-model-router&lt;/code&gt; if anything looks wrong. It prints every event hooked and every call it makes.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Code Mode makes Claude Code write JavaScript instead of calling tools one at a time</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Mon, 16 Nov 2026 15:54:12 +0000</pubDate>
      <link>https://dev.to/dani_avila7/code-mode-makes-claude-code-write-javascript-instead-of-calling-tools-one-at-a-time-2p46</link>
      <guid>https://dev.to/dani_avila7/code-mode-makes-claude-code-write-javascript-instead-of-calling-tools-one-at-a-time-2p46</guid>
      <description>&lt;p&gt;Claude Code 2.1.323 added Code Mode, an experimental opt-in where Claude stops asking for Bash, Read and Edit one at a time and instead writes a piece of JavaScript that makes those calls. Claude Code runs that code in your session. One script can fan out commands in parallel, feed one result into the next call, and hand back only the slice of output the model needs to read.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2121789088499933223-781" src="https://platform.twitter.com/embed/Tweet.html?id=2121789088499933223"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2121789088499933223-781');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2121789088499933223&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;I tried it with Fable 5.2 on a task that is all plumbing: find every Markdown link in the project that points at a local file that does not exist. It came back with 13 tool calls in 2 model turns and all 3 broken links. Without Code Mode that is 13 round trips, each one spending a turn of context and a wait.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2121789095168868770-541" src="https://platform.twitter.com/embed/Tweet.html?id=2121789095168868770"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2121789095168868770-541');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2121789095168868770&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  What changes when Code Mode is on
&lt;/h2&gt;

&lt;p&gt;In the conversation you see what you always saw: the edit diffs, the command output. You do not see the JavaScript. Press Ctrl+O to open the transcript viewer and each script is there as a collapsed Code row.&lt;/p&gt;

&lt;p&gt;The code looks like ordinary scripting. A shell call goes through a tagged template, an edit goes through a helper, and the last expression is what Claude reads back.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;files&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; $&lt;span class="sb"&gt;`rg -l "function retryPolicy" src`&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;text&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;edit&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;files&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;trim&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"\n"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"maxAttempts: 3"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"maxAttempts: 5"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That shift matters for debugging. The script is the unit of work now, so when a run surprises you, the script is the thing to read, not the tool list.&lt;/p&gt;

&lt;h2&gt;
  
  
  Turning it on, and the one place the variable has to go
&lt;/h2&gt;

&lt;p&gt;The variable is CLAUDE_CODE_EXPERIMENTAL_CODE_MODE set to 1. You can export it for one session, or put it in your user settings so every new session starts with it.&lt;/p&gt;

&lt;p&gt;I packaged the settings version as a &lt;a href="https://www.aitmpl.com/component/settings/environment/code-mode" rel="noopener noreferrer"&gt;Code Mode setting&lt;/a&gt; so you do not have to hand edit JSON:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest --setting environment/code-mode
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpmkfa3h7k44wo7ks70x5.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpmkfa3h7k44wo7ks70x5.jpeg" width="800" height="468"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2121789098994135137-106" src="https://platform.twitter.com/embed/Tweet.html?id=2121789098994135137"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2121789098994135137-106');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2121789098994135137&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;One thing to get right: Claude Code ignores this variable in project and local settings files. That is deliberate, so a repository you clone cannot switch Code Mode on for you. When the installer asks which settings file to merge into, pick User settings. Anything else installs cleanly and does nothing.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"CLAUDE_CODE_EXPERIMENTAL_CODE_MODE"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"1"&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first session asks you to confirm. After you accept, Claude Code restarts and prints &lt;code&gt;Code Mode (experimental) is on&lt;/code&gt; near the top. If you get &lt;code&gt;Code Mode (experimental) isn't available&lt;/code&gt; instead, the reason is on the same line and &lt;code&gt;/status&lt;/code&gt; repeats it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Permissions still run, but auto mode reviews the code
&lt;/h2&gt;

&lt;p&gt;Every tool call a script makes still goes through your deny and ask rules and your PreToolUse hooks. You do not need a new matcher for this: a Bash call from inside a script is matched by your Bash rules the same way it was before.&lt;/p&gt;

&lt;p&gt;Auto mode is where the shape changes. The classifier reviews each piece of code before it runs, not each tool call, and it does not see values computed at run time, such as a path taken out of search output. If there is a call you genuinely need to control, put it in a rule or a hook instead of leaning on that review.&lt;/p&gt;

&lt;p&gt;To check the JavaScript itself, write a PreToolUse hook that matches REPL, the tool that runs Claude’s code, and read it from &lt;code&gt;tool_input.code&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;One caveat the post did not mention: Bash rules match command text, so a denied command can still be written in a form your rule does not match. That is true with or without Code Mode. Sandboxing is the limit that does not depend on how the command is spelled.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you give up
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Resuming: a session started in Code Mode resumes only with Code Mode still on. With it off, &lt;code&gt;claude --continue&lt;/code&gt; and &lt;code&gt;claude --resume&lt;/code&gt; refuse the session.&lt;/li&gt;
&lt;li&gt;Time: code you wait on stops after ten minutes, including the time you spend at permission prompts. Background code gets up to two hours, and only if you started the session in auto or bypassPermissions mode.&lt;/li&gt;
&lt;li&gt;Surfaces: interactive terminal, native build. Not &lt;code&gt;-p&lt;/code&gt;, the Agent SDK, the VS Code panel, the Desktop app or cloud sessions, and not Remote Control, Claude in Chrome or computer use.&lt;/li&gt;
&lt;li&gt;Models: Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2.&lt;/li&gt;
&lt;li&gt;Accounts: a Claude subscription or Console account. Not through a gateway, Bedrock, Google Cloud or Microsoft Foundry, and not under the HIPAA configuration or zero data retention.&lt;/li&gt;
&lt;li&gt;Conflicts: it will not start with an API key helper, agent teams, OpenTelemetry export or DISABLE_TELEMETRY in place.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;For a team, the off switch is the same variable with the value 0 in managed settings. People already inside a Code Mode session stay there until they restart, and after that nobody can resume one. Managed settings can block it, they cannot turn it on.&lt;/p&gt;

&lt;p&gt;Is the trade worth it? For a repetitive edit across many files, a search whose output needs filtering before the model reads it, or a few independent commands that can run at once, yes. For a two call task you are adding a code generation step to something that was already one round trip. It is experimental, and Anthropic says it can change or be removed in any release, so do not build a workflow that depends on it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the settings live
&lt;/h2&gt;

&lt;p&gt;The Code Mode setting lives in the catalog next to the other things you would put in &lt;code&gt;~/.claude/settings.json&lt;/code&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/component/settings/environment/code-mode" rel="noopener noreferrer"&gt;Code Mode setting&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.aitmpl.com/component/settings/permissions/deny-sensitive-files" rel="noopener noreferrer"&gt;Deny Sensitive Files&lt;/a&gt;, the deny rules that keep applying to every call a script makes&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/settings/" rel="noopener noreferrer"&gt;The settings catalog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;Claude Code Templates&lt;/a&gt; for the rest&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftnc778f3oqren8k0wjaj.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftnc778f3oqren8k0wjaj.jpeg" width="800" height="468"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Senior Frontend SKILL makes Claude Code open your page in a real browser before it says done</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Wed, 11 Nov 2026 15:54:13 +0000</pubDate>
      <link>https://dev.to/dani_avila7/senior-frontend-skill-makes-claude-code-open-your-page-in-a-real-browser-before-it-says-done-26f3</link>
      <guid>https://dev.to/dani_avila7/senior-frontend-skill-makes-claude-code-open-your-page-in-a-real-browser-before-it-says-done-26f3</guid>
      <description>&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2120513853725294751-511" src="https://platform.twitter.com/embed/Tweet.html?id=2120513853725294751"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2120513853725294751-511');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2120513853725294751&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;A passing typecheck tells you nothing about whether the button can be reached with Tab, or whether a long name breaks the layout on a phone. That gap is the whole reason &lt;a href="https://www.aitmpl.com/component/skills/development/senior-frontend" rel="noopener noreferrer"&gt;senior-frontend&lt;/a&gt; exists. It is a skill in the Claude Code Templates catalog, 31K downloads, currently v2.0.0, and what it really does is refuse to let Claude call a UI task finished on the strength of a clean build.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the skill actually changes
&lt;/h2&gt;

&lt;p&gt;Most of the value is ordering. The skill makes Claude read before it writes: the framework and its version, the existing &lt;code&gt;Button&lt;/code&gt; and &lt;code&gt;Dialog&lt;/code&gt; components, the design tokens, the data layer, the npm scripts. In a project that already has tokens, a new hex value counts as a bug.&lt;/p&gt;

&lt;p&gt;Then it forces the states nobody asks for. Loading, empty, error, success, pending action, long and odd content, partial failure. Each one has a required outcome, not a suggestion: the empty state has to say why it is empty and what to do about it, the error has to say what failed and offer a way out, and one row’s failed action must not break the rest of the list.&lt;/p&gt;

&lt;p&gt;The build rules are platform first. &lt;code&gt;&amp;lt;button&amp;gt;&lt;/code&gt; for actions, &lt;code&gt;&amp;lt;a href&amp;gt;&lt;/code&gt; for navigation, never a &lt;code&gt;&amp;lt;div onClick&amp;gt;&lt;/code&gt;. Every input gets a visible label, the right &lt;code&gt;type&lt;/code&gt;, and &lt;code&gt;autocomplete&lt;/code&gt;. Icon-only buttons get an accessible name that says which thing they act on, so you do not end up with twelve buttons all called "Delete". Async results land in an &lt;code&gt;aria-live&lt;/code&gt; region.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install it and look at what lands
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2w0476boavv63hd4inlb.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2w0476boavv63hd4inlb.jpeg" width="800" height="442"&gt;&lt;/a&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest --skill development/senior-frontend
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That writes 6 files to &lt;code&gt;.claude/skills/senior-frontend/&lt;/code&gt;: the &lt;code&gt;SKILL.md&lt;/code&gt;, two reference docs (a UI review checklist and a set of React state and effect rules), a NOTICE file, and two scripts. Copy the folder to &lt;code&gt;~/.claude/skills/&lt;/code&gt; if you want it in every project instead of one.&lt;/p&gt;

&lt;p&gt;Only the name and description sit in your context until the skill is invoked. The body loads when it fires, and the reference docs load only when Claude reaches for them. Worth knowing, because the description is deliberately aggressive: it opens with “You MUST use this whenever you build, change or review user interface code, including small UI tweaks and when the user is in a hurry”. That is what makes it catch the one line CSS fix, and also what will make it fire more often than you might want. If that bothers you, &lt;code&gt;disable-model-invocation: true&lt;/code&gt; in the frontmatter leaves it available as &lt;code&gt;/senior-frontend&lt;/code&gt; and nothing else.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that costs you time
&lt;/h2&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2114401602450706524-137" src="https://platform.twitter.com/embed/Tweet.html?id=2114401602450706524"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2114401602450706524-137');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2114401602450706524&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;Step 6 of the skill is where the 3x execution time comes from. After typecheck, lint, test and build, Claude is told to start the dev server and run the bundled checker against the page it just built:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;node .claude/skills/senior-frontend/scripts/check_ui.mjs &lt;span class="s2"&gt;"http://localhost:5173/#/team"&lt;/span&gt; --out .ui-check
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The script loads the page at 375 px and 1280 px (those are the defaults, &lt;code&gt;--widths&lt;/code&gt; changes them), saves a full page screenshot at each width, and reports console errors, horizontal overflow, controls with no accessible name, inputs with no label, images with no alt or no explicit size, and clickable elements that are not interactive. At the widest width it presses Tab up to 40 times and lists every focus stop, flagging the ones with no visible focus indicator.&lt;/p&gt;

&lt;p&gt;One caveat the post leaves out: this step is not free. The script resolves Playwright from your project or a global install and exits with code 2 if it finds neither, so a project without Playwright gets the checklist and no browser pass. axe-core is optional in the same way: if it resolves from your project, WCAG 2.2 AA violations get folded into the report; if not, the report says so and moves on.&lt;/p&gt;

&lt;p&gt;The instruction after that one matters more than the script: look at the screenshots, do not just confirm they exist. Then force the error state, search for nothing, remove the last item, submit something invalid, and use only the keyboard.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is the trade worth it
&lt;/h2&gt;

&lt;p&gt;I measured it on Sonnet 5.5 and browser verification went from 0/8 without the skill to 10/10 with it, at roughly three times the execution time. That is one project and not a benchmark, but the direction is easy to believe: an agent that is told to open the page will open the page, and an agent that is not will tell you it builds.&lt;/p&gt;

&lt;p&gt;The honest version of the trade is that you are paying wall clock time to move a class of bugs from your review to the agent’s. If you are prototyping, that is a bad deal and you should skip the skill. If the thing you are building will be used with a keyboard, on a phone, or by someone who will hit the error state, it pays for itself the first time.&lt;/p&gt;

&lt;p&gt;The skill also ends every run with a report: what it built, which existing components and tokens it reused, what it verified and how, and what it could not verify. That last part is the one to read.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the skills live
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2550x320h1dl3zriue7k.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2550x320h1dl3zriue7k.jpeg" width="800" height="442"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2120365886880059433-594" src="https://platform.twitter.com/embed/Tweet.html?id=2120365886880059433"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2120365886880059433-594');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2120365886880059433&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Component page: &lt;a href="https://www.aitmpl.com/component/skills/development/senior-frontend" rel="noopener noreferrer"&gt;senior-frontend on aitmpl.com&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Related: &lt;a href="https://www.aitmpl.com/component/skills/development/systematic-debugging" rel="noopener noreferrer"&gt;systematic-debugging&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Full skills catalog: &lt;a href="https://www.aitmpl.com/skills/" rel="noopener noreferrer"&gt;aitmpl.com/skills&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Everything else: &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Claude Code reads AGENTS.md now, but only when CLAUDE.md is missing</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Sun, 08 Nov 2026 15:52:36 +0000</pubDate>
      <link>https://dev.to/dani_avila7/claude-code-reads-agentsmd-now-but-only-when-claudemd-is-missing-4lmh</link>
      <guid>https://dev.to/dani_avila7/claude-code-reads-agentsmd-now-but-only-when-claudemd-is-missing-4lmh</guid>
      <description>&lt;p&gt;Claude Code shipped AGENTS.md support on September 18, 2026. The changelog line for 2.1.277 is one sentence, and every word in it matters:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;Added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead; change it under “Project instructions” in /config (not yet on Bedrock, Vertex or Foundry)&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2101023103962718457-239" src="https://platform.twitter.com/embed/Tweet.html?id=2101023103962718457"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2101023103962718457-239');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101023103962718457&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;Read that again: &lt;strong&gt;in a project with no CLAUDE.md&lt;/strong&gt;. This is a fallback, not a merge. If your repo already has a CLAUDE.md, nothing changed for you. Claude Code keeps reading CLAUDE.md and keeps ignoring AGENTS.md, exactly as before.&lt;/p&gt;

&lt;h2&gt;
  
  
  What 2.1.277 actually changed
&lt;/h2&gt;

&lt;p&gt;Three things, and only three.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;When Claude Code finds no CLAUDE.md, it reads AGENTS.md instead of loading nothing.&lt;/li&gt;
&lt;li&gt;The behavior is togglable under &lt;strong&gt;Project instructions&lt;/strong&gt; in &lt;code&gt;/config&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;It is not available yet on Amazon Bedrock, Google Vertex AI, or Microsoft Foundry.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;There is no documented &lt;code&gt;settings.json&lt;/code&gt; key and no environment variable for the toggle. The only documented surface is the &lt;code&gt;/config&lt;/code&gt; menu. If you deploy Claude Code through managed settings, that matters: you cannot pin this behavior across a fleet today.&lt;/p&gt;

&lt;p&gt;One more thing worth knowing before you build on it. The memory documentation page has not caught up. At the time of writing it still says, in plain text, that Claude Code reads CLAUDE.md and not AGENTS.md. The changelog is ahead of the docs here, so trust the release notes and verify in your own session.&lt;/p&gt;

&lt;h2&gt;
  
  
  The fallback almost never fires in a real repo
&lt;/h2&gt;

&lt;p&gt;Claude Code does not look for one instructions file. It walks a tree.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/                     &amp;lt;- managed policy CLAUDE.md
~/.claude/CLAUDE.md   &amp;lt;- user instructions, every project
repo/
  CLAUDE.md           &amp;lt;- project instructions (or .claude/CLAUDE.md)
  CLAUDE.local.md     &amp;lt;- your gitignored personal layer
  services/api/
    CLAUDE.md         &amp;lt;- loads on demand, when Claude reads files here
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;At launch, Claude Code loads CLAUDE.md and CLAUDE.local.md from the working directory and every directory above it. Files in subdirectories are not loaded at launch; they are pulled in when Claude reads files in those subdirectories.&lt;/p&gt;

&lt;p&gt;Everything discovered is concatenated, not overridden. Order runs from the filesystem root down to your working directory, so the instructions closest to where you launched Claude are read last. Inside a single directory, &lt;code&gt;CLAUDE.local.md&lt;/code&gt; is appended after &lt;code&gt;CLAUDE.md&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Now put the AGENTS.md fallback next to that. It only has an opening when there is no CLAUDE.md at all in the project. In a repo that has been through &lt;code&gt;/init&lt;/code&gt; even once, that opening does not exist. Which is why, for most teams, the two older mechanisms are still the ones doing the work.&lt;/p&gt;

&lt;h2&gt;
  
  
  The import still beats the fallback
&lt;/h2&gt;

&lt;p&gt;If you keep AGENTS.md for Codex, Cursor, or anything else, the documented way to share it is a CLAUDE.md that imports it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;@AGENTS.md
&lt;span class="gu"&gt;## Claude Code&lt;/span&gt;
Use plan mode for changes under &lt;span class="sb"&gt;`src/billing/`&lt;/span&gt;.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Claude loads the imported file at session start, then appends whatever you wrote below it. You get one source of truth plus a Claude-specific section, and you are not relying on a fallback that disappears the moment someone commits a CLAUDE.md.&lt;/p&gt;

&lt;p&gt;The import rules are strict, and they bite:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Syntax is &lt;code&gt;@path/to/import&lt;/code&gt;. Relative paths resolve against the file containing the import, &lt;strong&gt;not&lt;/strong&gt; your working directory.&lt;/li&gt;
&lt;li&gt;Imports nest to a maximum depth of four hops.&lt;/li&gt;
&lt;li&gt;Import parsing skips Markdown code spans and fenced code blocks. Wrapping &lt;code&gt;@README&lt;/code&gt; in backticks keeps it literal; outside backticks it imports the file.&lt;/li&gt;
&lt;li&gt;An import in a project file that resolves outside the working directory is external. The first time Claude Code sees one, it shows an approval dialog. Decline once and those imports stay off, with no second prompt.&lt;/li&gt;
&lt;li&gt;Imports in user-scope files such as &lt;code&gt;~/.claude/CLAUDE.md&lt;/code&gt; load without that dialog, because you wrote them.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That last pair is the one that surprises people on a shared repo: your teammate approved the external import, you declined it, and now the two of you are running different instructions.&lt;/p&gt;

&lt;h2&gt;
  
  
  The symlink, and the one-time copy
&lt;/h2&gt;

&lt;p&gt;If you need no Claude-specific additions at all, a symlink is enough:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;ln&lt;/span&gt; &lt;span class="nt"&gt;-s&lt;/span&gt; AGENTS.md CLAUDE.md
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It prints nothing on success. On Windows this needs Administrator privileges or Developer Mode, so use the &lt;code&gt;@AGENTS.md&lt;/code&gt; import there instead.&lt;/p&gt;

&lt;p&gt;The third option is a one-time migration rather than a live link. &lt;code&gt;/import&lt;/code&gt; brings a supported coding agent configuration into Claude Code: it appends a copy of instruction files such as AGENTS.md to the matching CLAUDE.md, and carries over MCP servers, commands, subagents, and skills. It requires Claude Code v2.1.213 or later. Use it when you are leaving the other tool, not when you are running both.&lt;/p&gt;

&lt;p&gt;Related, and easy to miss: &lt;code&gt;/init&lt;/code&gt; reads Cursor rules and Copilot rules by default. Set &lt;code&gt;CLAUDE_CODE_NEW_INIT=1&lt;/code&gt; and it also reads AGENTS.md, &lt;code&gt;.devin/rules/&lt;/code&gt;, &lt;code&gt;.windsurf/rules/&lt;/code&gt; or &lt;code&gt;.windsurfrules&lt;/code&gt;, and &lt;code&gt;.clinerules&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check what actually loaded
&lt;/h2&gt;

&lt;p&gt;Every mechanism above is invisible until you verify it. Three tools, in order of effort:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/context   -&amp;gt; "Memory files" section lists what loaded this session
/memory    -&amp;gt; browse and open every CLAUDE.md and memory file
hook       -&amp;gt; InstructionsLoaded logs which files loaded, when, and why
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;/context&lt;/code&gt; is the fast check. If your file is not under &lt;strong&gt;Memory files&lt;/strong&gt;, Claude cannot see it, and no amount of rewording the instructions will help. The &lt;code&gt;InstructionsLoaded&lt;/code&gt; hook is the one to reach for when a path-scoped rule or a lazily loaded subdirectory file is the suspect.&lt;/p&gt;

&lt;p&gt;Two limits that quietly change behavior while you debug: Claude Code loads a CLAUDE.md up to 4 MiB in full and skips a larger file entirely, and adherence drops well before that, which is why the docs target under 200 lines per file. In a monorepo, &lt;code&gt;claudeMdExcludes&lt;/code&gt; skips ancestor files by glob against absolute paths, at any settings layer, with arrays merging across layers. A managed policy CLAUDE.md cannot be excluded.&lt;/p&gt;

&lt;h2&gt;
  
  
  Do this, in this order
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Run &lt;code&gt;claude --version&lt;/code&gt;. The fallback needs 2.1.277 or later, and it is not on Bedrock, Vertex, or Foundry.&lt;/li&gt;
&lt;li&gt;Check whether the repo has a CLAUDE.md. If it does, the new fallback will never fire, and steps 3 and 4 are your real options.&lt;/li&gt;
&lt;li&gt;Keep AGENTS.md as the source of truth and add a CLAUDE.md containing &lt;code&gt;@AGENTS.md&lt;/code&gt; plus any Claude-specific section.&lt;/li&gt;
&lt;li&gt;If you need nothing Claude-specific, use &lt;code&gt;ln -s AGENTS.md CLAUDE.md&lt;/code&gt; instead, unless you are on Windows without Developer Mode.&lt;/li&gt;
&lt;li&gt;Start a session and run &lt;code&gt;/context&lt;/code&gt;. Confirm the file appears under &lt;strong&gt;Memory files&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Open &lt;code&gt;/config&lt;/code&gt;, find &lt;strong&gt;Project instructions&lt;/strong&gt;, and set the AGENTS.md behavior you want before someone else on the team sets a different one.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Brainstorming gets a design doc approved before Claude Code builds anything</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Thu, 05 Nov 2026 15:54:23 +0000</pubDate>
      <link>https://dev.to/dani_avila7/brainstorming-gets-a-design-doc-approved-before-claude-code-builds-anything-4gm5</link>
      <guid>https://dev.to/dani_avila7/brainstorming-gets-a-design-doc-approved-before-claude-code-builds-anything-4gm5</guid>
      <description>&lt;p&gt;Claude Code is good at building, and it builds what it believes I asked for. That belief is usually thinner than it feels, and the gap is where I lose afternoons. &lt;a href="https://www.aitmpl.com/component/skill/development/brainstorming" rel="noopener noreferrer"&gt;Brainstorming&lt;/a&gt; is the skill I install to close it. It stops at the idea, asks until it understands what I actually want, and does not start building until I have approved a design.&lt;/p&gt;

&lt;p&gt;It is adapted from Jesse Vincent’s superpowers collection, it sits in the development category of the catalog, and it is at version 2.0.0.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2116718444364083536-331" src="https://platform.twitter.com/embed/Tweet.html?id=2116718444364083536"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2116718444364083536-331');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2116718444364083536&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  What Brainstorming does to a session
&lt;/h2&gt;

&lt;p&gt;It does not wait for the word “brainstorm”. Its description tells Claude to use it before creating features, adding functionality, modifying behavior or planning anything new, in software or out of it.&lt;/p&gt;

&lt;p&gt;Once it fires, the order is fixed:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;It reads first. In an existing project it goes through the code the change would touch, and it is told never to ask what the project already answers.&lt;/li&gt;
&lt;li&gt;One question per message. One sentence, no stacked questions, no menu of example answers attached. The red flags table in the skill is blunt about why a menu is worse than a question: options steer, so get them describing it first.&lt;/li&gt;
&lt;li&gt;It plays the design back in chunks of 200 to 300 words as it goes, marking which parts are mine and which are its own proposals.&lt;/li&gt;
&lt;li&gt;It sends the approaches in a message of their own: two or three real options with trade-offs, the weak ones dropped, its recommendation first with a reason.&lt;/li&gt;
&lt;li&gt;It writes the design doc and asks for approval.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;There is an escape hatch, and it is the reason a description that greedy is tolerable. If I never asked for brainstorming, it adds one line offering to just dive in instead. If I take it, the skill is done for that turn.&lt;/p&gt;

&lt;h2&gt;
  
  
  The gate, and what actually enforces it
&lt;/h2&gt;

&lt;p&gt;The approval gate is the part worth knowing, because it is wider than “do not write code”. The skill counts all of this as building:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;product code&lt;/li&gt;
&lt;li&gt;scaffolding&lt;/li&gt;
&lt;li&gt;installing dependencies&lt;/li&gt;
&lt;li&gt;creating projects&lt;/li&gt;
&lt;li&gt;invoking an implementation skill&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That last one closes the usual workaround, where the agent does not write the code itself, it just calls the skill that does. Reading, exploring, and a mockup or a spike I agreed to stay allowed, and keeping what a spike produced is treated as a new request that comes back through the gate.&lt;/p&gt;

&lt;p&gt;Here is the caveat the skill file does not give you. This gate is prompt text. Claude Code’s permissions documentation is explicit that rules are enforced by Claude Code and not by the model, and that instructions in a prompt or in CLAUDE.md shape what Claude tries to do rather than what it is allowed to do. Nothing in the runtime refuses a Write call because a design doc has not been approved. Two things follow:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Skill content loads once and is not re-read, so in a long session the gate is a memory rather than a check.&lt;/li&gt;
&lt;li&gt;If you want it actually enforced, that is a PreToolUse hook or a deny rule, not a skill.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I run it as a thinking aid, not as a control, and on those terms it earns the install.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install it, then decide how loudly it fires
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest &lt;span class="nt"&gt;--skill&lt;/span&gt; development/brainstorming
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That writes to &lt;code&gt;.claude/skills/brainstorming/&lt;/code&gt;: the SKILL.md plus a &lt;code&gt;references/builder-check.md&lt;/code&gt; that only gets read when it is needed. Only the frontmatter sits in context at session start, the body loads when the skill is invoked, and the reference file later still.&lt;/p&gt;

&lt;p&gt;Then read that description again, because the description is the trigger. “You MUST use this before any creative or build work”, with “skip it only for fully specified one-line tasks”, means the skill wants to fire on almost everything. On a repo where I am making small, already decided changes, that is noise.&lt;/p&gt;

&lt;p&gt;Two switches:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"skillOverrides"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"brainstorming"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"user-invocable-only"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;user-invocable-only&lt;/code&gt; keeps it out of automatic invocation and leaves &lt;code&gt;/brainstorming&lt;/code&gt; working when I want it. The &lt;code&gt;/skills&lt;/code&gt; menu writes this into &lt;code&gt;.claude/settings.local.json&lt;/code&gt;. The other values are on, name-only and off.&lt;/p&gt;

&lt;p&gt;The skill-side equivalent is &lt;code&gt;disable-model-invocation: true&lt;/code&gt; in the frontmatter, which also stops it being preloaded into a subagent.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the design doc lands
&lt;/h2&gt;

&lt;p&gt;After I pick an approach it sizes the work, and I can override the size. A small change gets a short design in the chat, a line or two per point. A project gets a written document, and in a repo it goes to &lt;code&gt;docs/plans/YYYY-MM-DD-&amp;lt;topic&amp;gt;-design.md&lt;/code&gt; unless the project already has its own convention. Both include an Approaches section: what was considered, what it trades off, which one I picked and why.&lt;/p&gt;

&lt;p&gt;For projects there is one more step I had not seen inside a skill before. Before the doc is handed over it runs a builder check: a fresh reader, a subagent or itself reading as a stranger, gets the document with the conversation stripped out and is asked for nothing except the questions a builder would still have to guess at, ranked. What comes back is sorted into already answered, minor (settle it and mark it as its own call), and mine. The last pile comes back in a single message. One round, then the doc is finished.&lt;/p&gt;

&lt;p&gt;One caveat there too: the SKILL.md hands off to a &lt;code&gt;writing-plans&lt;/code&gt; skill "if installed". There is no documented mechanism for one skill to depend on another, so that sentence is prompt text like the rest. If you do not have that skill, nothing fails and nothing warns you, you just get a numbered plan instead.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this sits above plan mode
&lt;/h2&gt;

&lt;p&gt;Plan mode is a permission mode. It holds back edits while a plan is written for work that has already been decided. Brainstorming runs upstream of that, on deciding what the work is.&lt;/p&gt;

&lt;p&gt;Claude Code did ship something closer to this once. I wrote about Ultraplan when it quietly disappeared, and my read then was that the idea was fine and the execution environment was the problem.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2085494482993106950-245" src="https://platform.twitter.com/embed/Tweet.html?id=2085494482993106950"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2085494482993106950-245');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2085494482993106950&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;Whether writing things down first actually pays is not settled either. The CMU and Google study on spec driven development I went through covered 266 repositories and 21 practitioners. Most specs carried implementation details, developers reported more reliable agents along with more planning and review overhead, and the researchers could not separate the benefit of the method from models simply getting better.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2114486060730237224-158" src="https://platform.twitter.com/embed/Tweet.html?id=2114486060730237224"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2114486060730237224-158');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2114486060730237224&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;What I keep from that is the part that does not depend on the document: the value is being forced to say what I want out loud, early. A skill that asks one question at a time is a cheap way to be forced.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the Skills live
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwjw3bq0579jot5wvxqu8.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fwjw3bq0579jot5wvxqu8.jpeg" width="800" height="746"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/component/skill/development/brainstorming" rel="noopener noreferrer"&gt;Brainstorming on aitmpl.com&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.aitmpl.com/component/skill/development/systematic-debugging" rel="noopener noreferrer"&gt;Systematic Debugging&lt;/a&gt;, the sibling skill for after it is built&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/skills" rel="noopener noreferrer"&gt;The rest of the skills catalog&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgp224r9cxwffg65ax2lb.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgp224r9cxwffg65ax2lb.jpeg" width="800" height="746"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Everything else is at &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Name your Claude Code sessions before you run more than one</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Tue, 03 Nov 2026 15:53:30 +0000</pubDate>
      <link>https://dev.to/dani_avila7/name-your-claude-code-sessions-before-you-run-more-than-one-2d3f</link>
      <guid>https://dev.to/dani_avila7/name-your-claude-code-sessions-before-you-run-more-than-one-2d3f</guid>
      <description>&lt;p&gt;you are looking at&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2095277618807316566-141" src="https://platform.twitter.com/embed/Tweet.html?id=2095277618807316566"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2095277618807316566-141');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2095277618807316566&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;The shorthand in that post is missing its arguments. Both flags take a value, so the real command is closer to this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude &lt;span class="nt"&gt;--name&lt;/span&gt; auth-refactor &lt;span class="nt"&gt;--worktree&lt;/span&gt; auth-refactor
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Once you run two or three sessions side by side, the thing that breaks first is not context or tokens. It is you, looking at four terminal panes and not knowing which one is editing which branch.&lt;/p&gt;

&lt;h2&gt;
  
  
  What -n and -w actually do
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;--name&lt;/code&gt; (short form &lt;code&gt;-n&lt;/code&gt;) sets a display name for the session. It shows up in the session picker, in the &lt;code&gt;/resume&lt;/code&gt; list, in the prompt bar, and in the &lt;code&gt;session_name&lt;/code&gt; field the status line receives. You can set it mid-session with &lt;code&gt;/rename&lt;/code&gt;. If two sessions end up with the same name, Claude Code renames one with a two word suffix, so you get something like &lt;code&gt;auth-refactor-graceful-unicorn&lt;/code&gt; instead of two identical entries.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;--worktree&lt;/code&gt; (short form &lt;code&gt;-w&lt;/code&gt;) takes a worktree name and runs the session inside a git worktree instead of your main checkout. It creates the worktree if it does not exist. It also accepts a PR reference, which is the form worth remembering:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude &lt;span class="nt"&gt;-n&lt;/span&gt; review-1234 &lt;span class="nt"&gt;-w&lt;/span&gt; &lt;span class="s1"&gt;'#1234'&lt;/span&gt;
claude &lt;span class="nt"&gt;-n&lt;/span&gt; hotfix &lt;span class="nt"&gt;-w&lt;/span&gt; https://github.com/org/repo/pull/1234
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Inside a worktree session, Claude Code enforces isolation. Edits to files in the main checkout, git commands that would redirect it, and shell commands that reach outside are blocked. That is the actual reason to use &lt;code&gt;-w&lt;/code&gt; rather than just opening a second terminal in the same directory.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the worktree lands, and on which branch
&lt;/h2&gt;

&lt;p&gt;New worktrees go to &lt;code&gt;.claude/worktrees/&amp;lt;name&amp;gt;/&lt;/code&gt; at your repository root. The branch is named &lt;code&gt;worktree-&amp;lt;name&amp;gt;&lt;/code&gt;, unless you branched from a PR, in which case it is &lt;code&gt;pr-&amp;lt;number&amp;gt;&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The default that catches people: a new worktree branches from the repository default branch on the remote, not from your current HEAD. That is the &lt;code&gt;fresh&lt;/code&gt; setting, and it is the documented default. If you want new worktrees to start from the work you have in front of you, change it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"worktree"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"baseRef"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"head"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The second gotcha is gitignored files. A fresh worktree has no &lt;code&gt;.env&lt;/code&gt;, so anything that reads it fails immediately. Add a &lt;code&gt;.worktreeinclude&lt;/code&gt; at the project root, next to &lt;code&gt;.git/&lt;/code&gt;, using gitignore syntax. Files are copied only when they both match a pattern and are gitignored:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;.env
.env.local
config/secrets.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Clean worktrees are removed when the session exits. Ones with uncommitted work stay, and you remove them yourself with &lt;code&gt;git worktree remove&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  /color is per session and does not persist
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;/color&lt;/code&gt; sets the prompt bar color. It accepts &lt;code&gt;red&lt;/code&gt;, &lt;code&gt;blue&lt;/code&gt;, &lt;code&gt;green&lt;/code&gt;, &lt;code&gt;yellow&lt;/code&gt;, &lt;code&gt;purple&lt;/code&gt;, &lt;code&gt;orange&lt;/code&gt;, &lt;code&gt;pink&lt;/code&gt;, &lt;code&gt;cyan&lt;/code&gt;, or &lt;code&gt;default&lt;/code&gt; to reset. With no argument it picks one at random.&lt;/p&gt;

&lt;p&gt;It is a session-scoped setting. Close the session and the color is gone, so it is a visual tag for the panes open right now, not part of your project config. Run it once after the session starts:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;/color red
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Put the name and the worktree in the status line
&lt;/h2&gt;

&lt;p&gt;The name and the worktree only help if you can see them without switching panes. The status line is where that goes. Point &lt;code&gt;statusLine&lt;/code&gt; at a script in any settings file, user, project, or local:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"statusLine"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"~/.claude/statusline.sh"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"padding"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The script gets a JSON object on stdin. The fields that matter here are &lt;code&gt;session_name&lt;/code&gt;, the &lt;code&gt;worktree&lt;/code&gt; object, and &lt;code&gt;context_window&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;#!/usr/bin/env bash&lt;/span&gt;
&lt;span class="nv"&gt;input&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;&lt;span class="nb"&gt;cat&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;
&lt;span class="nv"&gt;name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;jq &lt;span class="nt"&gt;-r&lt;/span&gt; &lt;span class="s1"&gt;'.session_name // "unnamed"'&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&amp;lt;&amp;lt;&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$input&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;
&lt;span class="nv"&gt;tree&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;jq &lt;span class="nt"&gt;-r&lt;/span&gt; &lt;span class="s1"&gt;'.worktree.name // "main checkout"'&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&amp;lt;&amp;lt;&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$input&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;
&lt;span class="nv"&gt;branch&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;jq &lt;span class="nt"&gt;-r&lt;/span&gt; &lt;span class="s1"&gt;'.worktree.branch // "-"'&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&amp;lt;&amp;lt;&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$input&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;
&lt;span class="nv"&gt;pct&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;jq &lt;span class="nt"&gt;-r&lt;/span&gt; &lt;span class="s1"&gt;'.context_window.used_percentage // 0'&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&amp;lt;&amp;lt;&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$input&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;
&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s | %s (%s) | ctx %s%%'&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$name&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$tree&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$branch&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$pct&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The fallbacks are not defensive padding. The &lt;code&gt;worktree&lt;/code&gt; object is present only while the session is actually in a worktree, and &lt;code&gt;session_name&lt;/code&gt; is absent when the session has neither a custom name nor an AI generated title. Without the defaults you get empty strings in half your panes.&lt;/p&gt;

&lt;p&gt;The full payload carries more than this: &lt;code&gt;model.display_name&lt;/code&gt;, &lt;code&gt;workspace.project_dir&lt;/code&gt;, &lt;code&gt;worktree.original_branch&lt;/code&gt;, &lt;code&gt;cost&lt;/code&gt;, &lt;code&gt;prompt_cache&lt;/code&gt;, &lt;code&gt;rate_limits&lt;/code&gt;, &lt;code&gt;vim.mode&lt;/code&gt;. If you would rather not write the script by hand, &lt;code&gt;/statusline&lt;/code&gt; takes a natural language description and generates one, then writes the settings entry for you.&lt;/p&gt;

&lt;p&gt;There is an installable version of the one from the video:&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2096292257322181033-993" src="https://platform.twitter.com/embed/Tweet.html?id=2096292257322181033"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2096292257322181033-993');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2096292257322181033&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Start here, in this order
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Set &lt;code&gt;worktree.baseRef&lt;/code&gt; to &lt;code&gt;head&lt;/code&gt; in your settings file if you want worktrees to branch from your current work rather than the remote default.&lt;/li&gt;
&lt;li&gt;Add a &lt;code&gt;.worktreeinclude&lt;/code&gt; at the repo root listing the gitignored files each worktree needs.&lt;/li&gt;
&lt;li&gt;Write the status line script and register it under &lt;code&gt;statusLine&lt;/code&gt;, or generate one with &lt;code&gt;/statusline&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Start each session with &lt;code&gt;claude -n &amp;lt;name&amp;gt; -w &amp;lt;name&amp;gt;&lt;/code&gt;, using the same string for both so the pane, the branch, and the directory all match.&lt;/li&gt;
&lt;li&gt;Run &lt;code&gt;/color &amp;lt;color&amp;gt;&lt;/code&gt; once inside each session.&lt;/li&gt;
&lt;li&gt;Confirm the status line shows the name and the branch before you start a second session, not after you have four.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Twitter&lt;/strong&gt; (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;LinkedIn&lt;/strong&gt; (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Youtube&lt;/strong&gt; (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Instagram&lt;/strong&gt; (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Superlogical MCP puts the Rex docs, changelog and Discord threads in your Claude Code session</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Mon, 02 Nov 2026 20:25:17 +0000</pubDate>
      <link>https://dev.to/dani_avila7/superlogical-mcp-puts-the-rex-docs-changelog-and-discord-threads-in-your-claude-code-session-37bi</link>
      <guid>https://dev.to/dani_avila7/superlogical-mcp-puts-the-rex-docs-changelog-and-discord-threads-in-your-claude-code-session-37bi</guid>
      <description>&lt;p&gt;I got early access to Rex, the terminal from Superlogical, and spent the week moving my sessions over from Ghostty. The part that is hard to write about is that Rex is still a beta: the macOS app and the Linux and FreeBSD servers are out to testers, parts of the docs sit behind a Discord login, and most of the real knowledge, the issues, the workarounds, the reasoning behind a decision, lives in archived threads in the Rex Beta Discord.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2115542930697908587-467" src="https://platform.twitter.com/embed/Tweet.html?id=2115542930697908587"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2115542930697908587-467');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2115542930697908587&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;That is a bad shape for an agent. You cannot point Claude Code at a docs site it cannot fully read, and you certainly cannot point it at a Discord server. So Superlogical shipped a community MCP for exactly this, and I added it to the catalog as &lt;a href="https://www.aitmpl.com/component/mcp/devtools/superlogical" rel="noopener noreferrer"&gt;superlogical&lt;/a&gt; today.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx5a7ykdc17p2xp20bevk.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx5a7ykdc17p2xp20bevk.jpeg" width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What the server actually serves
&lt;/h2&gt;

&lt;p&gt;Three things, all of them things you would otherwise go read by hand:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The Rex documentation, including the pages that are gated behind a tester login.&lt;/li&gt;
&lt;li&gt;The release notes for every Rex version, pulled from the &lt;code&gt;#changelog&lt;/code&gt; channel.&lt;/li&gt;
&lt;li&gt;Current and archived discussions from the Rex Beta Discord, including the Superlogical team own commentary on issues and design decisions.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The third one is the reason to install it. Documentation tells you what a flag does. The archived threads tell you why the flag exists and what broke before it did, and that is the context you are missing when an agent is doing the reading for you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Installing it and signing in
&lt;/h2&gt;

&lt;p&gt;It is a remote HTTP server, so there is nothing to run locally. Either command works:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# From the aitmpl catalog&lt;/span&gt;
npx claude-code-templates@latest &lt;span class="nt"&gt;--mcp&lt;/span&gt; devtools/superlogical
&lt;span class="c"&gt;# Or the command Superlogical documents&lt;/span&gt;
claude mcp add &lt;span class="nt"&gt;-s&lt;/span&gt; user &lt;span class="nt"&gt;--transport&lt;/span&gt; http superlogical https://www.superlogical.com/rex/community-mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;-s user&lt;/code&gt; puts it in &lt;code&gt;~/.claude.json&lt;/code&gt;, so it is available in every project on your machine and stays private to your account. That is the right scope here: this is a reference server about your tools, not a dependency of one repo.&lt;/p&gt;

&lt;p&gt;Then run &lt;code&gt;/mcp&lt;/code&gt; inside Claude Code and finish the login in the browser. The Claude Code docs are explicit that &lt;code&gt;/mcp&lt;/code&gt; is how you authenticate with remote servers that use OAuth 2.0, and this one uses Discord. &lt;code&gt;claude mcp list&lt;/code&gt; will show the server as needing authentication until you do, and connected afterwards.&lt;/p&gt;

&lt;p&gt;The gate is real and worth saying plainly: you need to be a member of the Rex Beta Discord. If you are not in the beta, installing this gets you a server you cannot sign in to. It is not a public docs mirror.&lt;/p&gt;

&lt;h2&gt;
  
  
  Name it superlogical, not rex
&lt;/h2&gt;

&lt;p&gt;This is the part none of my posts mentioned, and it is the one thing I would get wrong if I were installing it quickly.&lt;/p&gt;

&lt;p&gt;Superlogical recommends naming the server &lt;code&gt;superlogical&lt;/code&gt; rather than &lt;code&gt;rex&lt;/code&gt;, so it does not conflict with Rex skills, instructions, and the CLI. That sounds like a style note until you look at what Rex already installs. Running &lt;code&gt;rex skill setup&lt;/code&gt; writes a skill named &lt;code&gt;rex&lt;/code&gt; into &lt;code&gt;~/.claude/skills/rex&lt;/code&gt; and adds a hook to &lt;code&gt;~/.claude/settings.json&lt;/code&gt; so each session starts knowing it is inside Rex and which session and terminal it is in.&lt;/p&gt;

&lt;p&gt;So if you name the MCP server &lt;code&gt;rex&lt;/code&gt;, you now have a skill called &lt;code&gt;rex&lt;/code&gt;, a CLI called &lt;code&gt;rex&lt;/code&gt;, and a set of tools called &lt;code&gt;mcp__rex__*&lt;/code&gt;, all in the same session, all plausible answers when the model decides what to reach for. The install commands above already use the right name. Keep it.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2116166248820060390-945" src="https://platform.twitter.com/embed/Tweet.html?id=2116166248820060390"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2116166248820060390-945');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2116166248820060390&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;The naming also decides your permission rules, since tool names are built as &lt;code&gt;mcp__&amp;lt;server&amp;gt;__&amp;lt;tool&amp;gt;&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"permissions"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"allow"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"mcp__superlogical__*"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A glob is only allowed after the literal &lt;code&gt;mcp__superlogical__&lt;/code&gt; prefix. Do not write parentheses into an &lt;code&gt;mcp__&lt;/code&gt; rule: Claude Code skips any &lt;code&gt;mcp__&lt;/code&gt; rule containing them when it loads the settings file, and reports it in the invalid-settings dialog and in &lt;code&gt;claude doctor&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to expect once it is connected
&lt;/h2&gt;

&lt;p&gt;Searching documentation and a Discord archive returns a lot of text, and MCP output is capped. Claude Code warns when MCP tool output goes past 10,000 tokens and limits output to 25,000 tokens by default. &lt;code&gt;MAX_MCP_OUTPUT_TOKENS&lt;/code&gt; raises the limit; the warning threshold does not move. Separately, a successful text result longer than 50,000 characters is written to a file instead of being returned inline, and that rule ignores the variable entirely.&lt;/p&gt;

&lt;p&gt;In practice that means asking narrow questions. What changed in the last three Rex releases is a good question for this server. Summarize the Discord is how you spend your whole output budget on one call.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2115933086173028674-10" src="https://platform.twitter.com/embed/Tweet.html?id=2115933086173028674"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2115933086173028674-10');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2115933086173028674&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;p&gt;Is it worth it for a beta tool? For me, yes, and only because of the shape of the problem. I am setting up Tailscale to reach my sessions from my phone, and the answers to that kind of question are in threads, not in a reference page. Having the agent read them instead of me scrolling Discord is the entire value. If you are not in the beta, skip this one and come back when Rex opens up.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the MCPs live
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmethp7wholffh9mof2tt.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmethp7wholffh9mof2tt.jpeg" width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The component page: &lt;a href="https://www.aitmpl.com/component/mcp/devtools/superlogical" rel="noopener noreferrer"&gt;https://www.aitmpl.com/component/mcp/devtools/superlogical&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;A related one I use next to it, Firecrawl, for sites with no MCP of their own: &lt;a href="https://www.aitmpl.com/component/mcp/devtools/firecrawl" rel="noopener noreferrer"&gt;https://www.aitmpl.com/component/mcp/devtools/firecrawl&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The rest of the MCP catalog: &lt;a href="https://www.aitmpl.com/mcps" rel="noopener noreferrer"&gt;https://www.aitmpl.com/mcps&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Everything else: &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;https://www.aitmpl.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>mcp</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Stop correcting Claude in prose, rewind the session instead</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Fri, 30 Oct 2026 18:57:26 +0000</pubDate>
      <link>https://dev.to/dani_avila7/stop-correcting-claude-in-prose-rewind-the-session-instead-6fi</link>
      <guid>https://dev.to/dani_avila7/stop-correcting-claude-in-prose-rewind-the-session-instead-6fi</guid>
      <description>&lt;p&gt;When Claude makes a change you did not want, the reflex is to type a correction. “You shouldn’t have touched that file, undo it.” That message goes into the context window. Claude’s apology goes into the context window. The re-edit goes into the context window. And the bad version of the file is still sitting in the history, shaping everything that comes after. You paid tokens to make the session worse.&lt;/p&gt;

&lt;p&gt;There is a command for exactly this. &lt;code&gt;/rewind&lt;/code&gt; moves the session back to an earlier prompt and restores the files Claude edited along the way, so nothing about the mistake survives in the transcript.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2089701786374328691-713" src="https://platform.twitter.com/embed/Tweet.html?id=2089701786374328691"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2089701786374328691-713');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2089701786374328691&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  What /rewind actually restores
&lt;/h2&gt;

&lt;p&gt;Type &lt;code&gt;/rewind&lt;/code&gt;, or press Esc twice while the prompt input is empty, and you get a menu listing every prompt you sent in the session. Pick one and you are offered five actions:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Restore code and conversation&lt;/li&gt;
&lt;li&gt;Restore conversation, which rewinds the transcript and leaves the working tree alone&lt;/li&gt;
&lt;li&gt;Restore code, which reverts the files and keeps the transcript&lt;/li&gt;
&lt;li&gt;Summarize from here&lt;/li&gt;
&lt;li&gt;Summarize up to here&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The two code options only show up when that checkpoint has tracked file changes to revert. And if the prompt input has text in it, the first Esc clears the text instead of opening the menu. Press Up afterwards to pull it back out of input history.&lt;/p&gt;

&lt;h2&gt;
  
  
  What checkpointing does not track
&lt;/h2&gt;

&lt;p&gt;This is the part to internalize before you rely on it. Claude Code snapshots the file edits Claude makes through its own edit tools. That is the entire scope.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Edit / Write by Claude         tracked, /rewind restores it
Bash: rm, mv, cp               not tracked
Your own edits in the editor   not tracked
Background subagents           not tracked
Symlinks and hard links        skipped on restore
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The Bash exception is the one that bites. If Claude ran &lt;code&gt;rm old_config.yml&lt;/code&gt; as a shell command, &lt;code&gt;/rewind&lt;/code&gt; will not bring that file back. Same for &lt;code&gt;mv&lt;/code&gt; and &lt;code&gt;cp&lt;/code&gt;. Edits from a concurrent session are invisible too, and so are changes made by a background subagent. Foreground forked skills, the ones declared with &lt;code&gt;context: fork&lt;/code&gt; and &lt;code&gt;background: false&lt;/code&gt;, are the single subagent case that does get restored.&lt;/p&gt;

&lt;p&gt;Two more edges. A message you queue while Claude is still working joins the running turn and never gets its own checkpoint, so it will not appear in the menu at all. And restore skips symlinked and hard-linked paths, which is how dotfile managers and pnpm stores end up with “Restored the code, but skipped N files”.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;/debug
&lt;span class="c"&gt;# then read ~/.claude/debug/&amp;lt;session-id&amp;gt;.txt to see which paths were skipped and why&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;None of this touches git. Checkpointing is session-level recovery, not version history.&lt;/p&gt;

&lt;h2&gt;
  
  
  How long a checkpoint lives
&lt;/h2&gt;

&lt;p&gt;A checkpoint is created on every prompt that starts a turn. Claude Code keeps file snapshots for the 100 most recent checkpoints in a session, and a session’s snapshots are swept roughly 30 days after it last saved one. That window is &lt;code&gt;cleanupPeriodDays&lt;/code&gt; in &lt;code&gt;settings.json&lt;/code&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"cleanupPeriodDays"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;90&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"fileCheckpointingEnabled"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Rewinding to a checkpoint whose snapshots were already swept fails with &lt;code&gt;No files were restored&lt;/code&gt;. If you keep long-lived sessions around, raise the number before you need it, not after.&lt;/p&gt;

&lt;p&gt;Checkpoints are saved with the conversation, so resuming a session days later still gives you the full rewind menu. If you ran &lt;code&gt;/clear&lt;/code&gt; earlier in the same process, the menu puts a &lt;code&gt;/resume &amp;lt;session-id&amp;gt; (previous session)&lt;/code&gt; entry at the top so you can rewind past the clear. That entry needs v2.1.191 or later.&lt;/p&gt;

&lt;p&gt;To turn file snapshots off entirely, set &lt;code&gt;fileCheckpointingEnabled&lt;/code&gt; to false, or export &lt;code&gt;CLAUDE_CODE_DISABLE_CHECKPOINTS&lt;/code&gt; with any non-empty value.&lt;/p&gt;

&lt;h2&gt;
  
  
  It is not /clear, /compact or /branch
&lt;/h2&gt;

&lt;p&gt;These get mixed up constantly. Only one of them writes to your working tree.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;/rewind&lt;/code&gt; restores files and transcript to an earlier point. The only one that puts files back.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/clear&lt;/code&gt; starts an empty context and saves the old conversation so you can resume it. Files untouched.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/compact&lt;/code&gt; replaces the transcript with a summary in place, same session. Files untouched.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/resume&lt;/code&gt;, &lt;code&gt;--resume&lt;/code&gt; and &lt;code&gt;--continue&lt;/code&gt; load an existing session from disk. Files untouched.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/branch&lt;/code&gt; and &lt;code&gt;--fork-session&lt;/code&gt; copy the conversation up to now into a new session. Files untouched.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So when a change goes wrong, &lt;code&gt;/compact&lt;/code&gt; does nothing for you, and &lt;code&gt;/clear&lt;/code&gt; throws the good context away along with the bad. Rewind is the surgical one.&lt;/p&gt;

&lt;h2&gt;
  
  
  The order to do it in
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Notice the bad change before you start typing a correction.&lt;/li&gt;
&lt;li&gt;Clear the prompt input, then press Esc twice.&lt;/li&gt;
&lt;li&gt;Pick the prompt that came immediately before the mistake.&lt;/li&gt;
&lt;li&gt;Choose Restore code and conversation if Claude’s edits caused it, or Restore conversation if you only want the transcript cleaned up.&lt;/li&gt;
&lt;li&gt;Run &lt;code&gt;git status&lt;/code&gt; and check for anything Bash or a background subagent did, because rewind did not touch those.&lt;/li&gt;
&lt;li&gt;Rewrite the prompt with the guardrail that was missing, and send it again.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>productivity</category>
      <category>devtools</category>
    </item>
    <item>
      <title>Installs Claude Code Mods with one command</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Wed, 21 Oct 2026 15:03:00 +0000</pubDate>
      <link>https://dev.to/dani_avila7/installs-claude-code-mods-with-one-command-36h8</link>
      <guid>https://dev.to/dani_avila7/installs-claude-code-mods-with-one-command-36h8</guid>
      <description>&lt;p&gt;Mods let you hook into Claude Code while it works: block a command, rewrite a tool call, redact a secret from a result, or draw a live pane above the prompt. Until recently that meant writing the module yourself, packaging it as a plugin and publishing a marketplace. The Mods catalog on &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt; removes those steps: 38 mods, one command each.&lt;/p&gt;

&lt;h2&gt;
  
  
  What a Mod is
&lt;/h2&gt;

&lt;p&gt;A Mod is a Claude Code plugin whose behavior lives in a function-hooks module. A shell hook reacts to an event after the fact. A function hook wraps the event: your function receives the engine object &lt;code&gt;$&lt;/code&gt;, the event &lt;code&gt;e&lt;/code&gt;, and a &lt;code&gt;next&lt;/code&gt; continuation, and decides what happens.&lt;/p&gt;

&lt;p&gt;That gives three positions instead of one:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Observe.&lt;/strong&gt; Await &lt;code&gt;next(e)&lt;/code&gt; and inspect what came back.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rewrite.&lt;/strong&gt; Call &lt;code&gt;next({ ...e, input })&lt;/code&gt; with changed data, or change the result on the way out.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Answer.&lt;/strong&gt; Return &lt;code&gt;{ deny: "reason" }&lt;/code&gt; and never call &lt;code&gt;next&lt;/code&gt;. The tool does not run.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The module runs in a sandbox of its own, with no DOM and no Node. Everything outside it goes through &lt;code&gt;$&lt;/code&gt;: &lt;code&gt;$.ui&lt;/code&gt;, &lt;code&gt;$.http&lt;/code&gt;, &lt;code&gt;$.fs&lt;/code&gt;, &lt;code&gt;$.process&lt;/code&gt;, &lt;code&gt;$.state&lt;/code&gt;, &lt;code&gt;$.store&lt;/code&gt;, &lt;code&gt;$.session&lt;/code&gt;, &lt;code&gt;$.command&lt;/code&gt;, &lt;code&gt;$.tool&lt;/code&gt;, &lt;code&gt;$.model&lt;/code&gt; and &lt;code&gt;$.clock&lt;/code&gt;. That is what makes a mod more than a filter.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr97aiu38kdakq1gto39f.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fr97aiu38kdakq1gto39f.png" alt="image" width="800" height="492"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  It is a plugin, so there are two ways to get one
&lt;/h2&gt;

&lt;p&gt;Mods are distributed as plugins. The standard route is a plugin marketplace (a repo with &lt;code&gt;.claude-plugin/marketplace.json&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;/plugin marketplace add &amp;lt;repo&amp;gt;
/plugin &lt;span class="nb"&gt;install&lt;/span&gt; &amp;lt;mod&amp;gt;@&amp;lt;marketplace&amp;gt;
/reload-plugins
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;With aitmpl.com you skip that. Pick a mod on the site, copy the command, run it in your project:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest &lt;span class="nt"&gt;--mod&lt;/span&gt; security/secret-redactor
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The CLI downloads the whole plugin directory (&lt;code&gt;.claude-plugin/plugin.json&lt;/code&gt;, &lt;code&gt;hooks/&lt;/code&gt;, commands, types, tests) and writes it to &lt;code&gt;.claude/skills/secret-redactor/&lt;/code&gt;. Claude Code loads it on the next session as &lt;code&gt;secret-redactor@skills-dir&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4kfw06nixfe1f9i643qj.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4kfw06nixfe1f9i643qj.png" alt="image" width="799" height="434"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Two details that trip people up:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Trust.&lt;/strong&gt; A project’s &lt;code&gt;.claude/skills/&lt;/code&gt; is only read once you have accepted the workspace trust prompt. In a folder you never opened interactively (for example a headless &lt;code&gt;claude -p&lt;/code&gt; run) the mod does not load. Install it in &lt;code&gt;~/.claude/skills/&lt;/code&gt; if you want it everywhere.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Options.&lt;/strong&gt; Mod options are declared as &lt;code&gt;userConfig&lt;/code&gt; in &lt;code&gt;plugin.json&lt;/code&gt; and read from your &lt;strong&gt;user&lt;/strong&gt; settings, keyed by the full plugin id:
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"pluginConfigs"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"npm-to-pnpm-rewriter@skills-dir"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"options"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"manager"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"yarn"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Use the wrong key (&lt;code&gt;"npm-to-pnpm-rewriter"&lt;/code&gt;) and every option silently stays at its default.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is in the catalog
&lt;/h2&gt;

&lt;p&gt;Seven categories, 38 mods:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;security&lt;/strong&gt;: &lt;code&gt;secret-redactor&lt;/code&gt;, &lt;code&gt;block-destructive-commands&lt;/code&gt;, &lt;code&gt;protected-paths-guard&lt;/code&gt;, &lt;code&gt;large-edit-confirmation&lt;/code&gt;, &lt;code&gt;jev-auto-mode&lt;/code&gt;, &lt;code&gt;jev-guardrails&lt;/code&gt;, &lt;code&gt;jev-vercel-sandbox&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;productivity&lt;/strong&gt;: &lt;code&gt;aitmpl&lt;/code&gt; (a component browser inside Claude Code), &lt;code&gt;npm-to-pnpm-rewriter&lt;/code&gt;, &lt;code&gt;webfetch-cache&lt;/code&gt;, &lt;code&gt;jev-model-router&lt;/code&gt;, &lt;code&gt;jev-skill-suggestion&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;ui&lt;/strong&gt;: &lt;code&gt;agent-flow&lt;/code&gt;, &lt;code&gt;git-sidebar&lt;/code&gt;, &lt;code&gt;tool-timing-badge&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;observability&lt;/strong&gt;: &lt;code&gt;prompt-cache-control&lt;/code&gt;, &lt;code&gt;universal-audit-log&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;integrations&lt;/strong&gt;: &lt;code&gt;linear-tickets&lt;/code&gt;, &lt;code&gt;vercel-deploys&lt;/code&gt;, &lt;code&gt;neon-branch-per-session&lt;/code&gt;, &lt;code&gt;websearch-to-exa&lt;/code&gt;, &lt;code&gt;pi-agent-for-claude&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;enterprise&lt;/strong&gt;: &lt;code&gt;admin-capability-lockdown&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;games&lt;/strong&gt;: 15 of them above the prompt, from &lt;code&gt;snake&lt;/code&gt; and &lt;code&gt;tetris&lt;/code&gt; to &lt;code&gt;doom&lt;/code&gt;, &lt;code&gt;pacman&lt;/code&gt; and &lt;code&gt;tool-defense&lt;/code&gt;, where every enemy is one of Claude's real tool calls&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Terminal and Desktop
&lt;/h2&gt;

&lt;p&gt;Mods are not terminal-only. The official guide demos each one in both Claude Code in the terminal and Claude Code Desktop, and surfaces like panes and status lines are drawn through &lt;code&gt;$.ui.resolve(e)&lt;/code&gt;, which returns the constructors for whichever surface you are on. Same plugin, same install command, both places.&lt;/p&gt;

&lt;h2&gt;
  
  
  Requirements
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code 2.1.287 or later.&lt;/strong&gt; Per the official guide, mods are on by default; there is nothing to enable. (Earlier early-access builds needed &lt;code&gt;CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1&lt;/code&gt;.)&lt;/li&gt;
&lt;li&gt;The &lt;code&gt;$&lt;/code&gt; API is still young and can change between releases. The catalog's type definitions follow Claude Code 2.1.283 and every mod is typechecked against them in CI.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Debugging a mod that seems silent
&lt;/h2&gt;

&lt;p&gt;Before touching the code, check in this order:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Is the project trusted? (See above.)&lt;/li&gt;
&lt;li&gt;Are you looking in the right place? &lt;code&gt;$.ui.log&lt;/code&gt; goes to the transcript and to &lt;code&gt;~/.claude/debug/&amp;lt;session-id&amp;gt;.txt&lt;/code&gt;; in &lt;code&gt;claude -p&lt;/code&gt; it only reaches the debug file.&lt;/li&gt;
&lt;li&gt;Is &lt;code&gt;pluginConfigs&lt;/code&gt; keyed by &lt;code&gt;name@skills-dir&lt;/code&gt;?&lt;/li&gt;
&lt;li&gt;Does &lt;code&gt;claude --debug&lt;/code&gt; show &lt;code&gt;hooks module &amp;lt;name&amp;gt;@skills-dir loaded&lt;/code&gt;?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A note on model routers: a mod that rewrites &lt;code&gt;model&lt;/code&gt; or &lt;code&gt;effort&lt;/code&gt; at &lt;code&gt;turn.step&lt;/code&gt; changes that request only. The header, the status line and the effort box show your session's setting and will not move.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest &lt;span class="nt"&gt;--mod&lt;/span&gt; games/pacman
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Browse the catalog at &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt;, validate any installed mod with &lt;code&gt;claude plugin validate &amp;lt;dir&amp;gt;&lt;/code&gt;, and run its tests with &lt;code&gt;claude plugin test &amp;lt;dir&amp;gt;&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the Mods live
&lt;/h2&gt;

&lt;p&gt;Every mod has its own page with the install command, the options it accepts and what it hooks.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/component/mod/security/secret-redactor" rel="noopener noreferrer"&gt;Secret Redactor&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/component/mod/games/pacman" rel="noopener noreferrer"&gt;Pacman&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.aitmpl.com/mods" rel="noopener noreferrer"&gt;The full Mods catalog&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt; for the rest of the components: agents, commands, skills, hooks, MCPs and settings&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3l7d2jdhhmerzxj2blcl.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3l7d2jdhhmerzxj2blcl.jpeg" alt="image" width="800" height="408"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Ultra Config switches your whole Claude Code setup between five profiles mid-session</title>
      <dc:creator>Daniel (San) Ávila</dc:creator>
      <pubDate>Mon, 19 Oct 2026 15:00:58 +0000</pubDate>
      <link>https://dev.to/dani_avila7/ultra-config-switches-your-whole-claude-code-setup-between-five-profiles-mid-session-4199</link>
      <guid>https://dev.to/dani_avila7/ultra-config-switches-your-whole-claude-code-setup-between-five-profiles-mid-session-4199</guid>
      <description>&lt;p&gt;I keep changing the same handful of things in Claude Code depending on what I am doing. Cheap and quiet when I am grinding through a long refactor. The most capable model and maximum effort when the problem is actually hard. Everything asking first when I am working in someone else’s repo. Those are different values in different places, and I was editing them by hand every time.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.aitmpl.com/component/mod/productivity/ultra-config" rel="noopener noreferrer"&gt;Ultra Config&lt;/a&gt; is a Mod I built so that it is one press instead.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2111558707033018828-99" src="https://platform.twitter.com/embed/Tweet.html?id=2111558707033018828"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2111558707033018828-99');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2111558707033018828&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Five profiles, one press each
&lt;/h2&gt;

&lt;p&gt;Each profile is a list of settings with a reason attached to every one of them. This is what they set:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Token saver&lt;/strong&gt;: model &lt;code&gt;sonnet&lt;/code&gt;, effort &lt;code&gt;low&lt;/code&gt;, subagents on &lt;code&gt;haiku&lt;/code&gt;, Concise output, no prompt suggestions or recap, auto-compact on, small workflow fan-outs, the &lt;code&gt;ultracode&lt;/code&gt; keyword off, and MCP output capped at 10,000 tokens (the default is 25,000).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fast flow&lt;/strong&gt;: model &lt;code&gt;sonnet&lt;/code&gt;, effort &lt;code&gt;medium&lt;/code&gt;, edits auto-accepted, Concise output, prompt suggestions on, turn duration and the progress bar on, medium workflows.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Max quality&lt;/strong&gt;: the most capable model available to you, effort &lt;code&gt;xhigh&lt;/code&gt;, subagents keep whatever model they ask for, default output style, workflows on with large fan-outs and the &lt;code&gt;ultracode&lt;/code&gt; keyword, checkpoints and auto-compact on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Guarded&lt;/strong&gt;: every edit and command asks first, plan mode never acts on its own, &lt;code&gt;bypassPermissions&lt;/code&gt; disabled, Read deny rules for &lt;code&gt;.env&lt;/code&gt;, &lt;code&gt;secrets/&lt;/code&gt;, &lt;code&gt;~/.ssh&lt;/code&gt; and &lt;code&gt;~/.aws&lt;/code&gt;, checkpoints on, verbose output, &lt;code&gt;ultracode&lt;/code&gt; off.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Autopilot&lt;/strong&gt;: auto permission mode including during plan, effort &lt;code&gt;high&lt;/code&gt;, workflows on, recap on, push notifications when Claude decides something or needs you, auto-compact on.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The pane shows how much of a profile is already in force as a fraction, 3/11, and marks it active when all of it is. Open a profile and every line is either a change (old value struck through, new value, and why), something already set, or something skipped.&lt;/p&gt;

&lt;p&gt;Token saver is the one I reach for most, and it is worth noticing what it is: the same model split I posted about last week, without installing a separate setting for it.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2108687263043575923-831" src="https://platform.twitter.com/embed/Tweet.html?id=2108687263043575923"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2108687263043575923-831');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2108687263043575923&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Install it and open the pane
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx claude-code-templates@latest &lt;span class="nt"&gt;--mod&lt;/span&gt; productivity/ultra-config
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It needs Claude Code 2.1.287 or later, which is the version where Mods are on by default, and a trusted project folder.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftbd8sq1kgsqn5i735sil.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ftbd8sq1kgsqn5i735sil.jpeg" alt="Ultra Config" width="800" height="499"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Then:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;/ultra-config&lt;/code&gt; opens the pane&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/ultra-config &amp;lt;profile&amp;gt;&lt;/code&gt; opens it on one profile (&lt;code&gt;saver&lt;/code&gt;, &lt;code&gt;flow&lt;/code&gt;, &lt;code&gt;quality&lt;/code&gt;, &lt;code&gt;guarded&lt;/code&gt;, &lt;code&gt;autopilot&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/ultra-config apply &amp;lt;profile&amp;gt;&lt;/code&gt; activates one and prints what it changed and what it skipped&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/ultra-config undo&lt;/code&gt; reverts the last profile you activated, newest first&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/ultra-config close&lt;/code&gt; closes the pane&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Everything in the pane is a press, not a menu. There are no tabs and no arrow-key navigation, which matters because Claude Code already uses the left arrow for its own agents view.&lt;/p&gt;

&lt;p&gt;In the fullscreen layout the pane docks next to the transcript, which needs &lt;code&gt;/tui fullscreen&lt;/code&gt; and at least 110 columns. On the classic layout it sits above the prompt. In Claude Desktop it draws in the session's side panel.&lt;/p&gt;

&lt;p&gt;&lt;iframe class="tweet-embed" id="tweet-2111624442413756446-753" src="https://platform.twitter.com/embed/Tweet.html?id=2111624442413756446"&gt;
&lt;/iframe&gt;

  // Detect dark theme
  var iframe = document.getElementById('tweet-2111624442413756446-753');
  if (document.body.className.includes('dark-theme')) {
    iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2111624442413756446&amp;amp;theme=dark"
  }



&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the changes actually land
&lt;/h2&gt;

&lt;p&gt;This is the part worth understanding before you let something rewrite your configuration.&lt;/p&gt;

&lt;p&gt;Rows that exist in Claude Code’s own settings menu, like the model, the permission mode, the output style or auto-compact, are set through the same path the menu uses. Claude Code decides which file they go in, exactly as if you had changed them yourself.&lt;/p&gt;

&lt;p&gt;Effort is saved per model, the way &lt;code&gt;/effort&lt;/code&gt; saves it, under &lt;code&gt;modelSettings&lt;/code&gt; in &lt;code&gt;~/.claude/settings.json&lt;/code&gt;. That keying matters: since Claude Code 2.1.251, a top-level &lt;code&gt;effortLevel&lt;/code&gt; is ignored by Opus 5.5 and later models, so the Mod never writes one. A model id containing a dot, which is what you get on Bedrock and Vertex, cannot be saved this way, and the effort then applies to that one request only.&lt;/p&gt;

&lt;p&gt;Four keys have no menu row, so they are written straight into &lt;code&gt;~/.claude/settings.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"CLAUDE_CODE_SUBAGENT_MODEL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"haiku"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"MAX_MCP_OUTPUT_TOKENS"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"10000"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"permissions"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"disableBypassPermissionsMode"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"disable"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"deny"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Read(./.env)"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Read(./secrets/**)"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Read(~/.ssh/**)"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Before each of those writes the file is copied to &lt;code&gt;settings.json.ultra-config.bak&lt;/code&gt;, and a &lt;code&gt;settings.json&lt;/code&gt; that is not valid JSON is never overwritten: those steps are skipped and the pane says why. The last 20 activations are kept in the plugin's store, so &lt;code&gt;undo&lt;/code&gt; still works after a restart and puts back the values it changed, leaving alone any deny rule you already had.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check before you trust a profile
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Guarded does not stop a shell read.&lt;/strong&gt; The rules it adds are &lt;code&gt;Read()&lt;/code&gt; rules, and a Read deny rule blocks Claude's file tools, not a shell command that reads the same file. The Mod's own README says this plainly, which is more than most components do, but it is the thing people get wrong about deny rules.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The subagent model does not reach Explore and Plan.&lt;/strong&gt; &lt;code&gt;CLAUDE_CODE_SUBAGENT_MODEL&lt;/code&gt; on its own does not change the model the built-in Explore and Plan subagents run on. If you want Token saver's &lt;code&gt;haiku&lt;/code&gt; to apply to every subagent, you also need &lt;code&gt;CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1&lt;/code&gt;, which requires 2.1.257 or later.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two of those keys only take effect next session.&lt;/strong&gt; &lt;code&gt;CLAUDE_CODE_SUBAGENT_MODEL&lt;/code&gt; and &lt;code&gt;MAX_MCP_OUTPUT_TOKENS&lt;/code&gt; are read when Claude Code starts. The menu rows apply right away and the effort applies from your next message, but those two do not.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;disableBypassPermissionsMode locks you out too.&lt;/strong&gt; It is the string &lt;code&gt;"disable"&lt;/code&gt;, it works from any settings file, and while it is set Claude Code rejects &lt;code&gt;--dangerously-skip-permissions&lt;/code&gt; and ignores an agent definition asking for &lt;code&gt;bypassPermissions&lt;/code&gt;. Guarded writes it to your user settings, so &lt;code&gt;undo&lt;/code&gt; is how you get it back, not a flag.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Where the Mods live
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The component page: &lt;a href="https://www.aitmpl.com/component/mod/productivity/ultra-config" rel="noopener noreferrer"&gt;Ultra Config on aitmpl.com&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The model split on its own, without the rest of a profile: &lt;a href="https://www.aitmpl.com/component/setting/model/sonnet-opus-advisor-haiku-subagents" rel="noopener noreferrer"&gt;Sonnet Opus Advisor Haiku Subagents&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The rest of the catalog: &lt;a href="https://www.aitmpl.com" rel="noopener noreferrer"&gt;aitmpl.com&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4phs96086khvg2pbbuta.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4phs96086khvg2pbbuta.jpeg" alt="Ultra Config" width="800" height="499"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  You can follow me on my social media:
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Twitter (English): &lt;a href="https://x.com/dani_avila7" rel="noopener noreferrer"&gt;https://x.com/dani_avila7&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;LinkedIn (Spanish): &lt;a href="https://www.linkedin.com/in/daniel-avila-arias/" rel="noopener noreferrer"&gt;https://www.linkedin.com/in/daniel-avila-arias/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Youtube (Spanish): &lt;a href="https://www.youtube.com/@daniiielsan" rel="noopener noreferrer"&gt;https://www.youtube.com/@daniiielsan&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Instagram (Spanish): &lt;a href="https://www.instagram.com/danielsan_avila/" rel="noopener noreferrer"&gt;https://www.instagram.com/danielsan_avila/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
