<?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: Adrian Vale</title>
    <description>The latest articles on DEV Community by Adrian Vale (@adrian2016).</description>
    <link>https://dev.to/adrian2016</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%2F4217380%2F8c7345f1-03e7-4d90-84f6-7362a9275b77.png</url>
      <title>DEV Community: Adrian Vale</title>
      <link>https://dev.to/adrian2016</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/adrian2016"/>
    <language>en</language>
    <item>
      <title>What a Claude Code plugin runs before you type: we read 2,688 public plugins</title>
      <dc:creator>Adrian Vale</dc:creator>
      <pubDate>Fri, 06 Nov 2026 00:17:55 +0000</pubDate>
      <link>https://dev.to/adrian2016/what-a-claude-code-plugin-runs-before-you-type-we-read-2688-public-plugins-4agn</link>
      <guid>https://dev.to/adrian2016/what-a-claude-code-plugin-runs-before-you-type-we-read-2688-public-plugins-4agn</guid>
      <description>&lt;p&gt;Before you install a Claude Code plugin, open three things in its folder: &lt;code&gt;hooks/hooks.json&lt;/code&gt;, &lt;code&gt;.mcp.json&lt;/code&gt; and &lt;code&gt;bin/&lt;/code&gt;. Look for &lt;code&gt;SessionStart&lt;/code&gt; hooks, and for MCP servers started with &lt;code&gt;npx&lt;/code&gt; or &lt;code&gt;uvx&lt;/code&gt; but no version number.&lt;/p&gt;

&lt;p&gt;The reason is simple. A plugin can run code before you type anything. The docs say it directly: "A Claude Code plugin you install can execute arbitrary code on your machine with your user privileges." Four parts of a plugin start on their own:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;SessionStart hooks&lt;/strong&gt; run shell commands when a session starts. Their plain-text output is added to Claude's context, which Claude "can see and act on".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;MCP servers&lt;/strong&gt; declared with a command are processes Claude Code starts on your machine for an enabled plugin.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Monitors&lt;/strong&gt; are shell commands that run in the background for the whole session and send their output to Claude.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;bin/&lt;/code&gt;&lt;/strong&gt; is added to the &lt;code&gt;PATH&lt;/code&gt; of the shell Claude runs commands in.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Claude Code runs all of these outside the sandbox, and installing a plugin also enables it.&lt;/p&gt;

&lt;p&gt;So how many public plugins do this? We read every plugin in 1,961 active public repos.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check before you install a plugin
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Read &lt;code&gt;hooks/hooks.json&lt;/code&gt;, starting with &lt;code&gt;SessionStart&lt;/code&gt; and &lt;code&gt;Setup&lt;/code&gt;.&lt;/strong&gt; These run when a session starts, and their output becomes context for Claude. Read the scripts they call, not just the command line, since most of them run a file from the plugin folder.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read &lt;code&gt;.mcp.json&lt;/code&gt;.&lt;/strong&gt; A server with &lt;code&gt;command&lt;/code&gt; is a process on your machine. If it runs &lt;code&gt;npx some-package&lt;/code&gt; or &lt;code&gt;uvx some-package&lt;/code&gt; with no version, the package manager resolves a version at start. A new release can then reach your machine without any change to the plugin you reviewed. Pinned means a version after the name: &lt;code&gt;@1.4.2&lt;/code&gt; for npx, as in &lt;code&gt;npx -y some-package&lt;/code&gt; plus &lt;code&gt;@1.4.2&lt;/code&gt;, or &lt;code&gt;==1.4.2&lt;/code&gt; for uvx, as in &lt;code&gt;some-package==1.4.2&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;List everything before installing.&lt;/strong&gt; Clone the repo and run &lt;code&gt;claude --plugin-dir &amp;lt;plugin folder&amp;gt; plugin details &amp;lt;plugin name&amp;gt;&lt;/code&gt;. It reads the plugin's files without starting a session and prints its hooks with their events, MCP and LSP servers, skills and agents.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decide on updates.&lt;/strong&gt; With auto-update on for a marketplace, the docs note, "the files you reviewed can change on disk". Turn it off for marketplaces you don't control, or prefer ones that pin each plugin to a commit, as Anthropic's community marketplace does for nearly every entry.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If the plugin ships a mod, read our &lt;a href="https://agentcraft.dev/en/blog/claude-code-mods-review-2026-11/" rel="noopener noreferrer"&gt;review of 891 public mods&lt;/a&gt; for what to check there. We don't repeat it here.&lt;/p&gt;

&lt;h2&gt;
  
  
  How we tested
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find plugins.&lt;/strong&gt; GitHub code search finds 34,157 plugin manifests (&lt;code&gt;filename:plugin.json path:.claude-plugin&lt;/code&gt;). Search returns at most 1,000 results per query, so we split the search into 40 file-size slices. We ran them in alternating order under a 30-minute cap and finished 23, which cover every size range. That gave 17,965 unique manifests, about 53% of the total, from 17,928 repos.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Filter repos.&lt;/strong&gt; At least 5 stars, pushed to since October 7, not a fork, not archived: 1,964 repos. We read 1,961.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read every plugin in those repos.&lt;/strong&gt; We read each plugin's manifest, &lt;code&gt;hooks/hooks.json&lt;/code&gt; and the &lt;code&gt;hooks&lt;/code&gt; key, &lt;code&gt;.mcp.json&lt;/code&gt; and the &lt;code&gt;mcpServers&lt;/code&gt; key, &lt;code&gt;.lsp.json&lt;/code&gt;, &lt;code&gt;monitors/monitors.json&lt;/code&gt; and &lt;code&gt;bin/&lt;/code&gt;, plus the scripts their commands name. All files come from the same commit: 2,688 plugins, read as text. Nothing was installed or run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Weigh by repo.&lt;/strong&gt; Most repos hold one plugin; the largest holds 117. The main number is the share of repos with at least one such plugin. The other two are plugins capped at 10 per repo, and all plugins. They come out close.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What the plugin does&lt;/th&gt;
&lt;th&gt;Repos (of 1,961)&lt;/th&gt;
&lt;th&gt;Plugins, max 10 per repo&lt;/th&gt;
&lt;th&gt;All plugins&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Runs something when a session starts (SessionStart or Setup hook, local MCP server, or monitor)&lt;/td&gt;
&lt;td&gt;30.4%&lt;/td&gt;
&lt;td&gt;27.4%&lt;/td&gt;
&lt;td&gt;24.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SessionStart or Setup shell hook&lt;/td&gt;
&lt;td&gt;16.0%&lt;/td&gt;
&lt;td&gt;14.3%&lt;/td&gt;
&lt;td&gt;12.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;…whose script writes files&lt;/td&gt;
&lt;td&gt;6.0%&lt;/td&gt;
&lt;td&gt;5.1%&lt;/td&gt;
&lt;td&gt;4.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;…whose script goes online&lt;/td&gt;
&lt;td&gt;1.5%&lt;/td&gt;
&lt;td&gt;1.2%&lt;/td&gt;
&lt;td&gt;1.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Any shell hook&lt;/td&gt;
&lt;td&gt;22.2%&lt;/td&gt;
&lt;td&gt;20.1%&lt;/td&gt;
&lt;td&gt;18.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;UserPromptSubmit hook (sees each prompt, output goes to Claude)&lt;/td&gt;
&lt;td&gt;8.1%&lt;/td&gt;
&lt;td&gt;7.0%&lt;/td&gt;
&lt;td&gt;6.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;PreToolUse hook&lt;/td&gt;
&lt;td&gt;11.8%&lt;/td&gt;
&lt;td&gt;10.5%&lt;/td&gt;
&lt;td&gt;9.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Declares MCP servers&lt;/td&gt;
&lt;td&gt;26.3%&lt;/td&gt;
&lt;td&gt;24.1%&lt;/td&gt;
&lt;td&gt;21.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Local MCP server (a process on your machine)&lt;/td&gt;
&lt;td&gt;18.4%&lt;/td&gt;
&lt;td&gt;16.4%&lt;/td&gt;
&lt;td&gt;14.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Remote MCP server (a URL)&lt;/td&gt;
&lt;td&gt;8.6%&lt;/td&gt;
&lt;td&gt;8.2%&lt;/td&gt;
&lt;td&gt;7.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MCP server started by npx or bunx&lt;/td&gt;
&lt;td&gt;5.9%&lt;/td&gt;
&lt;td&gt;5.5%&lt;/td&gt;
&lt;td&gt;4.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;…with no version&lt;/td&gt;
&lt;td&gt;3.9%&lt;/td&gt;
&lt;td&gt;3.5%&lt;/td&gt;
&lt;td&gt;3.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MCP server started by uvx with no version&lt;/td&gt;
&lt;td&gt;1.4%&lt;/td&gt;
&lt;td&gt;1.2%&lt;/td&gt;
&lt;td&gt;1.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Any package-manager-started MCP server with no version&lt;/td&gt;
&lt;td&gt;5.3%&lt;/td&gt;
&lt;td&gt;4.7%&lt;/td&gt;
&lt;td&gt;4.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Passes an environment variable named like a secret to an MCP server&lt;/td&gt;
&lt;td&gt;2.4%&lt;/td&gt;
&lt;td&gt;2.1%&lt;/td&gt;
&lt;td&gt;1.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;bin/&lt;/code&gt; folder on Claude's PATH&lt;/td&gt;
&lt;td&gt;9.5%&lt;/td&gt;
&lt;td&gt;8.0%&lt;/td&gt;
&lt;td&gt;7.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ships a mod (see the mods review)&lt;/td&gt;
&lt;td&gt;5.3%&lt;/td&gt;
&lt;td&gt;4.9%&lt;/td&gt;
&lt;td&gt;4.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Monitors&lt;/td&gt;
&lt;td&gt;0.4%&lt;/td&gt;
&lt;td&gt;0.3%&lt;/td&gt;
&lt;td&gt;0.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sets &lt;code&gt;defaultEnabled: false&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;0.1%&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Quick guess: of the MCP servers these plugins start with &lt;code&gt;npx&lt;/code&gt; or &lt;code&gt;bunx&lt;/code&gt;, how many name a version? About one in three. Of 171 such servers, 112 (65%) name none: they use &lt;code&gt;@latest&lt;/code&gt;, a bare package name, or a GitHub reference. 149 (87%) pass &lt;code&gt;-y&lt;/code&gt;, which skips npm's install prompt. Across all 945 MCP servers, the most common kinds are remote URLs (445), local scripts in the plugin (176) and &lt;code&gt;npx&lt;/code&gt; (164).&lt;/p&gt;

&lt;p&gt;A few examples, without names:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Session start.&lt;/strong&gt; Most SessionStart hooks run a script from the plugin folder with &lt;code&gt;node&lt;/code&gt;, &lt;code&gt;bash&lt;/code&gt; or &lt;code&gt;python3&lt;/code&gt;, with file names like &lt;code&gt;session-start&lt;/code&gt;, &lt;code&gt;resume-hint&lt;/code&gt; or &lt;code&gt;context-inject&lt;/code&gt;. Of the 29 repos whose session-start script goes online, 26 check for updates against GitHub or npm, and several talk to their own service on &lt;code&gt;localhost&lt;/code&gt;. Around ten mention telemetry or analytics; we did not check what they send.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your API key.&lt;/strong&gt; One rule-checking hook calls Anthropic's API with your &lt;code&gt;ANTHROPIC_API_KEY&lt;/code&gt; and sends a slice of the conversation to judge it. Its comments say so. It spends your quota and sends your conversation, which you may want to know before installing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Credential paths.&lt;/strong&gt; Hooks that name &lt;code&gt;~/.ssh&lt;/code&gt; or &lt;code&gt;.env&lt;/code&gt; were mostly guards that list those paths in order to block them.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of the plugins our rules flagged looked malicious when we read them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to ignore
&lt;/h2&gt;

&lt;p&gt;The number of components. A plugin with twenty hooks can be a careful guard, and a plugin with one SessionStart script and an unpinned &lt;code&gt;npx&lt;/code&gt; server is the one to read closely. Look at what starts by itself and what it fetches.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A sample.&lt;/strong&gt; About 53% of manifests, from active repos with 5 or more stars, as of November 6, 2026. This is what public plugin repos contain, not what people actually install.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Static only.&lt;/strong&gt; We did not run hooks, so we don't know what context they add or what an online script sends.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unpinned is a risk, not a measurement of harm.&lt;/strong&gt; Whether a new version reaches you depends on the package manager's cache. Without a version, nothing in the plugin has to change for different code to run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Mods are counted, not analyzed.&lt;/strong&gt; See the mods review for those.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Reproduce it
&lt;/h2&gt;

&lt;p&gt;The counts above, in all three weightings, are in &lt;a href="https://agentcraft.dev/data/plugin-startup-2026-11.json" rel="noopener noreferrer"&gt;plugin-startup-2026-11.json&lt;/a&gt;. Repo and plugin names are left out. To check a plugin yourself, run &lt;code&gt;claude --plugin-dir &amp;lt;plugin folder&amp;gt; plugin details &amp;lt;plugin name&amp;gt;&lt;/code&gt; and read the files it lists.&lt;/p&gt;

&lt;p&gt;For skills you choose to install yourself, the &lt;a href="https://agentcraft.dev/en/skills/" rel="noopener noreferrer"&gt;AgentCraft skills directory&lt;/a&gt; lists 5,101 Claude Code skills, and every community skill is reviewed before it's listed.&lt;/p&gt;

&lt;p&gt;The full post is on AgentCraft: &lt;a href="https://agentcraft.dev/en/blog/claude-code-plugin-startup-2026-11/" rel="noopener noreferrer"&gt;https://agentcraft.dev/en/blog/claude-code-plugin-startup-2026-11/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Disclosure: I'm Adrian Vale and I help run AgentCraft.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>security</category>
      <category>mcp</category>
      <category>ai</category>
    </item>
    <item>
      <title>Before you open Claude Code in a cloned repo, check its .claude/skills folder</title>
      <dc:creator>Adrian Vale</dc:creator>
      <pubDate>Thu, 05 Nov 2026 22:21:51 +0000</pubDate>
      <link>https://dev.to/adrian2016/before-you-open-claude-code-in-a-cloned-repo-check-its-claudeskills-folder-1o0e</link>
      <guid>https://dev.to/adrian2016/before-you-open-claude-code-in-a-cloned-repo-check-its-claudeskills-folder-1o0e</guid>
      <description>&lt;p&gt;Before you open Claude Code in a repo you cloned, look inside its &lt;code&gt;.claude/skills/&lt;/code&gt; folder. Read the &lt;code&gt;allowed-tools&lt;/code&gt; line of each skill, search for &lt;code&gt;!`&lt;/code&gt;, and compare the folder names with the commands you plan to type.&lt;/p&gt;

&lt;p&gt;The reason is simple. A project skill comes with the repo, and three things about it don't depend on whether you trust the folder:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Pre-approved tools.&lt;/strong&gt; Whatever a skill lists in &lt;code&gt;allowed-tools&lt;/code&gt; runs without a permission prompt during the turn that uses it. The docs put it plainly: "Workspace trust never gates a skill's &lt;code&gt;allowed-tools&lt;/code&gt; in any session."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Commands on load.&lt;/strong&gt; A line like &lt;code&gt;!`git status`&lt;/code&gt; runs while the skill loads, before Claude reads it, with no prompt when your permission rules or the skill's &lt;code&gt;allowed-tools&lt;/code&gt; allow it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Same names.&lt;/strong&gt; A project skill named like a built-in command or bundled skill replaces it in a local terminal session. The docs' own example: "A project &lt;code&gt;code-review&lt;/code&gt; skill replaces &lt;code&gt;/code-review&lt;/code&gt;."&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And Claude can use most of these skills on its own. 98% of the repos we read ship a project skill without &lt;code&gt;disable-model-invocation&lt;/code&gt;, so Claude may load it whenever a request matches its description.&lt;/p&gt;

&lt;p&gt;So how common are these in practice? We read every project skill in 2,427 active public repos.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check before you open a cloned repo
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;List the skills.&lt;/strong&gt; &lt;code&gt;ls .claude/skills/&lt;/code&gt; at the repo root, plus any nested &lt;code&gt;.claude/skills/&lt;/code&gt; folders. If a folder is named like a command you use, such as &lt;code&gt;review&lt;/code&gt;, &lt;code&gt;code-review&lt;/code&gt;, &lt;code&gt;security-review&lt;/code&gt;, &lt;code&gt;verify&lt;/code&gt; or &lt;code&gt;debug&lt;/code&gt;, read that skill first. Typing the command will run the repo's version.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read &lt;code&gt;allowed-tools&lt;/code&gt;.&lt;/strong&gt; A bare &lt;code&gt;Bash&lt;/code&gt; means any shell command runs without asking during that turn. &lt;code&gt;Bash(git status *)&lt;/code&gt; allows only that command. Also look for &lt;code&gt;Write&lt;/code&gt;, &lt;code&gt;Edit&lt;/code&gt; and &lt;code&gt;WebFetch&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Search for &lt;code&gt;!`&lt;/code&gt;.&lt;/strong&gt; Each one is a command that runs when the skill loads. Most just read &lt;code&gt;git&lt;/code&gt; state; read the ones that do more.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;For a one-off look, start without project content.&lt;/strong&gt; For headless runs the docs suggest &lt;code&gt;claude -p --bare&lt;/code&gt;, which reads no hooks, skills, plugins or &lt;code&gt;.mcp.json&lt;/code&gt; from the project. &lt;code&gt;--setting-sources user&lt;/code&gt; skips the project's settings files and &lt;code&gt;.mcp.json&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  How we tested
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find project skills.&lt;/strong&gt; GitHub code search reports 506,880 files matching &lt;code&gt;filename:SKILL.md path:.claude/skills&lt;/code&gt;. Search returns at most 1,000 results per query, so we split the search into 58 file-size slices and sampled 41 of them. The 22 smallest slices ran in order, and every other slice ran after that, so every size range is represented. That gave 31,272 unique files, about 6% of the total, from 16,304 repos.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Filter repos.&lt;/strong&gt; At least 5 stars, pushed to since October 7, not a fork, not archived: 2,435 repos. We read 2,427.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read every project skill in those repos.&lt;/strong&gt; From each repo's file tree, every SKILL.md under &lt;code&gt;.claude/skills/&lt;/code&gt;, plus the scripts each one names: 41,142 skills. We read them as text. Nothing was installed or run.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Weigh by repo.&lt;/strong&gt; One repo holds 3,132 project skills and the median repo holds 7, so we report three numbers. The main one is the share of repos with at least one such skill, because that is what you meet when you clone. We also give the share of skills with at most 10 per repo, and the share of all 41,142 for comparison.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What the repo ships&lt;/th&gt;
&lt;th&gt;Repos (of 2,427)&lt;/th&gt;
&lt;th&gt;Skills, max 10 per repo&lt;/th&gt;
&lt;th&gt;All skills&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;A project skill Claude can invoke on its own&lt;/td&gt;
&lt;td&gt;97.9%&lt;/td&gt;
&lt;td&gt;94.3%&lt;/td&gt;
&lt;td&gt;95.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill with &lt;code&gt;allowed-tools&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;22.3%&lt;/td&gt;
&lt;td&gt;12.5%&lt;/td&gt;
&lt;td&gt;10.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill that pre-approves any Bash command&lt;/td&gt;
&lt;td&gt;12.6%&lt;/td&gt;
&lt;td&gt;6.4%&lt;/td&gt;
&lt;td&gt;5.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill that pre-approves listed Bash commands only&lt;/td&gt;
&lt;td&gt;10.4%&lt;/td&gt;
&lt;td&gt;3.9%&lt;/td&gt;
&lt;td&gt;2.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill that pre-approves Write or Edit&lt;/td&gt;
&lt;td&gt;11.5%&lt;/td&gt;
&lt;td&gt;5.0%&lt;/td&gt;
&lt;td&gt;4.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill that runs a command on load (&lt;code&gt;!`&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;2.8%&lt;/td&gt;
&lt;td&gt;0.8%&lt;/td&gt;
&lt;td&gt;0.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill that does both: &lt;code&gt;!`&lt;/code&gt; and any Bash&lt;/td&gt;
&lt;td&gt;0.5%&lt;/td&gt;
&lt;td&gt;0.1%&lt;/td&gt;
&lt;td&gt;0.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill that runs a bundled script&lt;/td&gt;
&lt;td&gt;29.5%&lt;/td&gt;
&lt;td&gt;7.4%&lt;/td&gt;
&lt;td&gt;7.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A skill whose script goes online&lt;/td&gt;
&lt;td&gt;10.9%&lt;/td&gt;
&lt;td&gt;1.6%&lt;/td&gt;
&lt;td&gt;1.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A root-level skill named like a built-in command or bundled skill&lt;/td&gt;
&lt;td&gt;16.4%&lt;/td&gt;
&lt;td&gt;2.1%&lt;/td&gt;
&lt;td&gt;1.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;A root-level skill named &lt;code&gt;review&lt;/code&gt;, &lt;code&gt;code-review&lt;/code&gt; or &lt;code&gt;security-review&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;6.3%&lt;/td&gt;
&lt;td&gt;0.7%&lt;/td&gt;
&lt;td&gt;0.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Quick guess: in how many of these repos does typing a common review command run something the repo's author wrote? 153 repos, about one in sixteen, have a root-level skill named &lt;code&gt;review&lt;/code&gt;, &lt;code&gt;code-review&lt;/code&gt; or &lt;code&gt;security-review&lt;/code&gt;. Across all 399 repos with a clash, the most common names are &lt;code&gt;verify&lt;/code&gt;, &lt;code&gt;review&lt;/code&gt;, &lt;code&gt;code-review&lt;/code&gt;, &lt;code&gt;debug&lt;/code&gt; and &lt;code&gt;plan&lt;/code&gt;. These look like ordinary names that happen to match. All 15 of the &lt;code&gt;review&lt;/code&gt;, &lt;code&gt;code-review&lt;/code&gt; and &lt;code&gt;security-review&lt;/code&gt; skills we sampled are genuine review workflows. They are still the repo author's review, with the author's rules, and some pre-approve Bash. The point is the mechanism: the command you type runs the repo's skill.&lt;/p&gt;

&lt;p&gt;The commands inside &lt;code&gt;!`&lt;/code&gt; are mostly harmless reads: &lt;code&gt;git&lt;/code&gt; (144 uses), &lt;code&gt;bash&lt;/code&gt; (79), &lt;code&gt;gh&lt;/code&gt; (36), &lt;code&gt;grep&lt;/code&gt;, &lt;code&gt;find&lt;/code&gt;, &lt;code&gt;cat&lt;/code&gt; and &lt;code&gt;ls&lt;/code&gt;. Five call &lt;code&gt;curl&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Most &lt;code&gt;allowed-tools: Bash&lt;/code&gt; lines look like authors saving themselves prompts in their own workflow. That is reasonable in your own repo. In someone else's repo, it means their skill runs shell commands in your session without asking, for the turn that uses it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hidden characters
&lt;/h2&gt;

&lt;p&gt;Claude Code started refusing skills whose folder names contain hidden, direction-changing or blank-like characters in 2.1.299 (October 14), and fixed other invisible-character problems in later releases. So we also looked for hidden text in the same 41,142 files:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What we looked for&lt;/th&gt;
&lt;th&gt;Repos&lt;/th&gt;
&lt;th&gt;Skills&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Zero-width characters&lt;/td&gt;
&lt;td&gt;26&lt;/td&gt;
&lt;td&gt;26&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Direction-changing characters&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Unicode tag characters (invisible text)&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Folder names the 2.1.299 rule would refuse&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;HTML comments over 40 characters, hidden when the file is rendered&lt;/td&gt;
&lt;td&gt;344&lt;/td&gt;
&lt;td&gt;2,087&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Such comments with "ignore previous" or "secretly" style wording&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;td&gt;1&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The zero-width characters we read were harmless: emoji joiners and escapes inside nested code blocks. The one comment with hiding words is a prompt-injection example inside a security testing guide.&lt;/p&gt;

&lt;p&gt;The one file with Unicode tag characters is different. It hides an instruction that a person reading the file cannot see, in a skill that also pre-approves Bash. The command it asks for is harmless, and it looks like a deliberate proof of concept. It shows the technique exists in a public repo today. We are not naming the repo. To check a repo for it, search its SKILL.md files for characters in the U+E0000 to U+E007F range.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to ignore
&lt;/h2&gt;

&lt;p&gt;The count of skills in a repo. A repo with hundreds of project skills can be harmless, and one skill with a bare &lt;code&gt;Bash&lt;/code&gt; and a familiar name is worth more of your attention. Read the folder names and the frontmatter.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A sample.&lt;/strong&gt; About 6% of matching files, from active repos with 5 or more stars, as of November 6, 2026. Code search returns results by relevance within each slice, so this is not a random sample.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Static only.&lt;/strong&gt; We read text with pattern rules and ran nothing. Rules miss commands built at runtime and can match commands quoted in documentation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Capabilities, not intent.&lt;/strong&gt; Almost everything here is authors configuring their own workflow. We counted what a skill can do in your session, and we did not judge why it does it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Name clashes count folder names.&lt;/strong&gt; The docs describe the replacement in local terminal sessions; we did not test every session type.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Reproduce it
&lt;/h2&gt;

&lt;p&gt;The counts above, in all three weightings, are in &lt;a href="https://agentcraft.dev/data/project-skills-2026-11.json" rel="noopener noreferrer"&gt;project-skills-2026-11.json&lt;/a&gt;. Repo and skill names are left out. To check a repo yourself, run the four checks above before you start Claude Code in it.&lt;/p&gt;

&lt;p&gt;For skills you choose to install yourself, the &lt;a href="https://agentcraft.dev/en/skills/" rel="noopener noreferrer"&gt;AgentCraft skills directory&lt;/a&gt; lists 5,101 Claude Code skills, and every community skill is reviewed before it's listed.&lt;/p&gt;

&lt;p&gt;The full post is on AgentCraft: &lt;a href="https://agentcraft.dev/en/blog/claude-code-project-skills-2026-11/" rel="noopener noreferrer"&gt;https://agentcraft.dev/en/blog/claude-code-project-skills-2026-11/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Disclosure: I'm Adrian Vale and I help run AgentCraft.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>security</category>
      <category>ai</category>
      <category>git</category>
    </item>
    <item>
      <title>One in five Claude Code skills runs a bundled script. We read what those scripts do</title>
      <dc:creator>Adrian Vale</dc:creator>
      <pubDate>Thu, 05 Nov 2026 19:48:41 +0000</pubDate>
      <link>https://dev.to/adrian2016/one-in-five-claude-code-skills-runs-a-bundled-script-we-read-what-those-scripts-do-34ag</link>
      <guid>https://dev.to/adrian2016/one-in-five-claude-code-skills-runs-a-bundled-script-we-read-what-those-scripts-do-34ag</guid>
      <description>&lt;p&gt;Before you install a Claude Code skill, open its folder and look for a &lt;code&gt;scripts/&lt;/code&gt; directory. If the SKILL.md tells Claude to run a file from it, read that file, and read the skill's &lt;code&gt;allowed-tools&lt;/code&gt; line too.&lt;/p&gt;

&lt;p&gt;The reason is simple. A skill is more than instructions. In 979 of the 5,036 community skills we read from our directory, about one in five, the SKILL.md tells Claude to run a script that ships with the skill. That script runs on your machine with your permissions. And if the skill's &lt;code&gt;allowed-tools&lt;/code&gt; lists &lt;code&gt;Bash&lt;/code&gt;, Claude can run it, and any other shell command, without asking you during that turn.&lt;/p&gt;

&lt;p&gt;Quick guess: how many of those scripts reach the network? 272 skills, about one in twenty. Many of them call the API the skill is about, and 93 of them also read an environment variable named like a secret, such as an API key.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three checks before you install a skill
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find the scripts the SKILL.md names, and search them.&lt;/strong&gt; Look for &lt;code&gt;curl&lt;/code&gt;, &lt;code&gt;requests&lt;/code&gt;, &lt;code&gt;fetch&lt;/code&gt; and &lt;code&gt;wget&lt;/code&gt; (network), &lt;code&gt;rm -rf&lt;/code&gt; and &lt;code&gt;shutil.rmtree&lt;/code&gt; (recursive deletes), &lt;code&gt;pip install&lt;/code&gt; and &lt;code&gt;npm install&lt;/code&gt; run from the script, and environment variables with &lt;code&gt;KEY&lt;/code&gt;, &lt;code&gt;TOKEN&lt;/code&gt; or &lt;code&gt;SECRET&lt;/code&gt; in the name. Each one is fine when it matches what the skill says it does.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read &lt;code&gt;allowed-tools&lt;/code&gt;.&lt;/strong&gt; The listed tools run without a permission prompt during the turn that invokes the skill. &lt;code&gt;Bash(git status *)&lt;/code&gt; allows that one command. A bare &lt;code&gt;Bash&lt;/code&gt; allows any command. The docs also note that workspace trust does not gate this field, so a project skill applies it even in a folder you have never trusted.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Search the SKILL.md for &lt;code&gt;!`&lt;/code&gt;.&lt;/strong&gt; A line like &lt;code&gt;!`git diff HEAD`&lt;/code&gt; runs a shell command when the skill loads, before Claude reads the skill, and pastes the output in. It is a useful feature, and it is also code that runs before anyone looks at it.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  How we tested
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Skills.&lt;/strong&gt; All 5,082 community skills in our directory, the ones that passed our review. The directory's other 19 are Anthropic's official skills, which we left out. We could read 5,036; the rest were in deleted repos or could not be fetched. We also read the 317 skills our November review dropped, to compare.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Files.&lt;/strong&gt; One GitHub tree per repo, then the SKILL.md and its scripts from that same commit. A script is a &lt;code&gt;.py&lt;/code&gt;, &lt;code&gt;.sh&lt;/code&gt;, &lt;code&gt;.js&lt;/code&gt;, &lt;code&gt;.ts&lt;/code&gt;, &lt;code&gt;.mjs&lt;/code&gt;, &lt;code&gt;.rb&lt;/code&gt;, &lt;code&gt;.pl&lt;/code&gt; or &lt;code&gt;.ps1&lt;/code&gt; file in the skill's folder. We skipped folders that belong to another skill, plus &lt;code&gt;node_modules&lt;/code&gt;, tests, fixtures and build output. We went at most three levels deep.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Runs a bundled script.&lt;/strong&gt; We count a skill here when its SKILL.md names one of those files. Everything we say about what scripts do is counted on these named scripts only.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reading.&lt;/strong&gt; Pattern rules on the source with comments removed. Nothing was installed or run. We checked samples by hand and fixed four rules that over-counted, for example printed "pip install X" hints that never run.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;Out of 5,036 skills in the directory:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What the skill does&lt;/th&gt;
&lt;th&gt;Skills&lt;/th&gt;
&lt;th&gt;Share&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Ships script files&lt;/td&gt;
&lt;td&gt;1,175&lt;/td&gt;
&lt;td&gt;23.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;SKILL.md tells Claude to run one of them&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;979&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;19.4%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script writes or deletes files&lt;/td&gt;
&lt;td&gt;691&lt;/td&gt;
&lt;td&gt;13.7%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script starts other programs&lt;/td&gt;
&lt;td&gt;342&lt;/td&gt;
&lt;td&gt;6.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script goes online&lt;/td&gt;
&lt;td&gt;272&lt;/td&gt;
&lt;td&gt;5.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script goes online and writes files&lt;/td&gt;
&lt;td&gt;236&lt;/td&gt;
&lt;td&gt;4.7%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script deletes folders recursively&lt;/td&gt;
&lt;td&gt;122&lt;/td&gt;
&lt;td&gt;2.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script reads an environment variable named like a secret&lt;/td&gt;
&lt;td&gt;119&lt;/td&gt;
&lt;td&gt;2.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script reads a secret-named variable and goes online&lt;/td&gt;
&lt;td&gt;93&lt;/td&gt;
&lt;td&gt;1.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script names a credential file (&lt;code&gt;.env&lt;/code&gt;, &lt;code&gt;~/.ssh&lt;/code&gt;, &lt;code&gt;~/.aws/credentials&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;57&lt;/td&gt;
&lt;td&gt;1.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script installs packages itself&lt;/td&gt;
&lt;td&gt;43&lt;/td&gt;
&lt;td&gt;0.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The script uses &lt;code&gt;sudo&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;td&gt;0.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SKILL.md sets &lt;code&gt;allowed-tools&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;563&lt;/td&gt;
&lt;td&gt;11.2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;allowed-tools&lt;/code&gt; pre-approves Write or Edit&lt;/td&gt;
&lt;td&gt;249&lt;/td&gt;
&lt;td&gt;4.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;allowed-tools&lt;/code&gt; pre-approves Bash for any command&lt;/td&gt;
&lt;td&gt;233&lt;/td&gt;
&lt;td&gt;4.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;allowed-tools&lt;/code&gt; pre-approves Bash for listed commands only&lt;/td&gt;
&lt;td&gt;151&lt;/td&gt;
&lt;td&gt;3.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Runs a bundled script and pre-approves Bash for any command&lt;/td&gt;
&lt;td&gt;50&lt;/td&gt;
&lt;td&gt;1.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SKILL.md runs a command on load with &lt;code&gt;!`&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;td&gt;0.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SKILL.md registers hooks that last the whole session&lt;/td&gt;
&lt;td&gt;13&lt;/td&gt;
&lt;td&gt;0.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A few examples, without names:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Network.&lt;/strong&gt; A website audit skill fetches the live pages it scores. A graph database skill calls the vendor's REST API to create and run agents. Both go online because that is the job.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Secrets.&lt;/strong&gt; An SEO checklist skill loads its own API keys from &lt;code&gt;~/.config/&amp;lt;tool&amp;gt;/.env&lt;/code&gt;. That is why it names a &lt;code&gt;.env&lt;/code&gt; file. It is also the line to read, since the same code could read any file you can.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deletes.&lt;/strong&gt; A music video skill removes a project folder with &lt;code&gt;shutil.rmtree&lt;/code&gt;. Check what path a recursive delete gets.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Installs.&lt;/strong&gt; A slide deck skill's PPTX export script runs &lt;code&gt;npm install&lt;/code&gt; inside the skill's own folder.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bash for any command.&lt;/strong&gt; An experiment loop skill pre-approves Read, Write, Edit, Bash and Agent, and runs its own tree script. Inside that turn, none of those tools asks you first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;On load.&lt;/strong&gt; A memory recall skill uses &lt;code&gt;!`&lt;/code&gt; to run a script from its plugin folder as soon as it loads.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Many of the 979 are ordinary: a scaffolder that creates the folders it needs, a deploy script that cleans up its temp folder. The point of the checks is to see that the script matches the skill.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the dropped skills tell us
&lt;/h2&gt;

&lt;p&gt;Our November review dropped 317 of the 783 new skills it read. We read those 317 the same way and matched each one to its published drop reason.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Why the review dropped it&lt;/th&gt;
&lt;th&gt;Skills&lt;/th&gt;
&lt;th&gt;Runs a bundled script&lt;/th&gt;
&lt;th&gt;Script goes online&lt;/th&gt;
&lt;th&gt;Script writes files&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Evades detection or platform rules&lt;/td&gt;
&lt;td&gt;35&lt;/td&gt;
&lt;td&gt;48.6%&lt;/td&gt;
&lt;td&gt;20.0%&lt;/td&gt;
&lt;td&gt;42.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Unsafe&lt;/td&gt;
&lt;td&gt;41&lt;/td&gt;
&lt;td&gt;39.0%&lt;/td&gt;
&lt;td&gt;14.6%&lt;/td&gt;
&lt;td&gt;29.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Paid API or ad funnel&lt;/td&gt;
&lt;td&gt;65&lt;/td&gt;
&lt;td&gt;21.5%&lt;/td&gt;
&lt;td&gt;16.9%&lt;/td&gt;
&lt;td&gt;12.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Only useful to its author or one project&lt;/td&gt;
&lt;td&gt;46&lt;/td&gt;
&lt;td&gt;13.0%&lt;/td&gt;
&lt;td&gt;6.5%&lt;/td&gt;
&lt;td&gt;10.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Placeholder, template, demo or test fixture&lt;/td&gt;
&lt;td&gt;108&lt;/td&gt;
&lt;td&gt;6.5%&lt;/td&gt;
&lt;td&gt;0%&lt;/td&gt;
&lt;td&gt;0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Kept in the directory&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;5,036&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;19.4%&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;5.4%&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;13.7%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Skills dropped as unsafe run a bundled script 2.0 times as often as the skills we kept (39.0% against 19.4%), and skills dropped for evading platform rules 2.5 times as often (48.6%). Their scripts go online 2.7 and 3.7 times as often (14.6% and 20.0% against 5.4%). Templates and demos almost never ship scripts, which pulls the dropped group's overall average down to 20.2%, close to the kept 19.4%.&lt;/p&gt;

&lt;p&gt;So a script is a normal part of a skill. In the two groups the review dropped for safety and rule breaking, scripts show up 2 to 2.5 times as often as in the skills we kept, and those scripts go online 2.7 to 3.7 times as often. Those are the skills where reading the script pays off most.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to ignore
&lt;/h2&gt;

&lt;p&gt;The file count. A skill with forty helper scripts can be harmless, and a skill with one can send your keys somewhere. Read the scripts the SKILL.md actually names.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Static only.&lt;/strong&gt; A named script may not run every time. Pattern rules miss calls built at runtime and can match a command inside a string. Network counts include some browser code that loads local files, and credential paths include security scanners that list those paths as patterns to block.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One snapshot.&lt;/strong&gt; Skills entered the directory over time. We read every script at its repo's current commit on November 6, 2026, so some authors have changed things since their skill was listed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not cause and effect.&lt;/strong&gt; The dropped skills are the ones our review flagged, not a random sample. The comparison shows which problems tend to come with scripts. It does not show that scripts get skills dropped.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Capabilities, not intent.&lt;/strong&gt; We did not judge whether any script misuses what it does.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Reproduce it
&lt;/h2&gt;

&lt;p&gt;The counts above, for kept skills, dropped skills and each drop reason, are in &lt;a href="https://agentcraft.dev/data/skill-scripts-2026-11.json" rel="noopener noreferrer"&gt;skill-scripts-2026-11.json&lt;/a&gt;. Repo and skill names are left out. To check one skill yourself, open its folder, find the file names its SKILL.md mentions, and search them for the patterns in the first check above.&lt;/p&gt;

&lt;p&gt;The skills we kept, with install commands, are in the &lt;a href="https://agentcraft.dev/en/skills/" rel="noopener noreferrer"&gt;AgentCraft skills directory&lt;/a&gt;: 5,101 Claude Code skills, and every community skill is reviewed before it's listed.&lt;/p&gt;

&lt;p&gt;The full post is on AgentCraft: &lt;a href="https://agentcraft.dev/en/blog/claude-code-skill-scripts-2026-11/" rel="noopener noreferrer"&gt;https://agentcraft.dev/en/blog/claude-code-skill-scripts-2026-11/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Disclosure: I'm Adrian Vale and I help run AgentCraft.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>security</category>
      <category>agents</category>
    </item>
    <item>
      <title>We read 891 public Claude Code mods before installing any. Here is what they can do</title>
      <dc:creator>Adrian Vale</dc:creator>
      <pubDate>Thu, 05 Nov 2026 19:01:58 +0000</pubDate>
      <link>https://dev.to/adrian2016/we-read-891-public-claude-code-mods-before-installing-any-here-is-what-they-can-do-1alf</link>
      <guid>https://dev.to/adrian2016/we-read-891-public-claude-code-mods-before-installing-any-here-is-what-they-can-do-1alf</guid>
      <description>&lt;p&gt;Before you install a Claude Code mod, run one command on its folder and read two lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude plugin validate ./some-mod
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;code&gt;calls:&lt;/code&gt; line tells you what the mod can reach: files, programs, the network, your environment variables. The &lt;code&gt;hooks:&lt;/code&gt; line tells you what it gets to see: every tool call, every prompt, the stored conversation. The command reads the code without running it, so it is safe to use on a mod you don't trust yet.&lt;/p&gt;

&lt;p&gt;The reason is simple. Mods arrived in Claude Code 2.1.287 on October 1, and they are not sandboxed. A mod runs with your permissions, and a process it starts runs outside the sandbox even when you have sandboxing on. The official docs say this plainly, and they suggest installing mods only from authors and marketplaces you trust.&lt;/p&gt;

&lt;p&gt;So what do the mods people are publishing actually ask for? We cloned 891 of them from GitHub and read what &lt;code&gt;claude plugin validate&lt;/code&gt; reports for each one. Most of them reach further than a status line needs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three checks before you install a mod
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Read the &lt;code&gt;calls:&lt;/code&gt; line.&lt;/strong&gt; Look for &lt;code&gt;$.http.fetch&lt;/code&gt; (network), &lt;code&gt;$.process.run&lt;/code&gt; or &lt;code&gt;$.process.spawn&lt;/code&gt; (starts programs), &lt;code&gt;$.fs.write&lt;/code&gt; (writes files), &lt;code&gt;$.env.get&lt;/code&gt; (reads environment variables, where API keys often live), &lt;code&gt;$.prompt.submit&lt;/code&gt; (sends prompts as you) and &lt;code&gt;$.model.complete&lt;/code&gt; (spends your plan or API key). Each of these is fine when it matches what the mod says it does. A token counter that goes online deserves a second look.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read the &lt;code&gt;hooks:&lt;/code&gt; line.&lt;/strong&gt; &lt;code&gt;tool.call&lt;/code&gt; without a filter in braces means the mod sees and can change every tool call. &lt;code&gt;prompt.submit&lt;/code&gt; means it sees every prompt you send. &lt;code&gt;session.append&lt;/code&gt; means it can rewrite rows of the conversation before they are stored. &lt;code&gt;tool.check&lt;/code&gt; means it can answer a permission question before you are asked.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;If it hooks &lt;code&gt;tool.check&lt;/code&gt;, find where it returns &lt;code&gt;allow&lt;/code&gt;.&lt;/strong&gt; That is the code that approves tool calls for you. Check what it lets through: its own tool only, read-only calls, or something wider.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  How we tested
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find repos.&lt;/strong&gt; GitHub code search for the mod file layout (a &lt;code&gt;hooks/hooks.json&lt;/code&gt; whose &lt;code&gt;modules&lt;/code&gt; list names a code file), plus repo searches for Claude Code mods: 1,781 repos.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Filter.&lt;/strong&gt; Not a fork, not archived, pushed to since October 7, and at least 5 stars: 311 repos. 303 of them cloned.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Find mods.&lt;/strong&gt; A mod counts when its &lt;code&gt;hooks.json&lt;/code&gt; points at a file that exports &lt;code&gt;register&lt;/code&gt;, the definition in the mods reference. That gave 915 mods. Identical files were counted once, keeping the most starred copy: 902.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validate.&lt;/strong&gt; We ran &lt;code&gt;claude plugin validate --json&lt;/code&gt; (Claude Code 2.1.314) on each one. We ran it with an empty home folder, so none of our own settings, hooks or plugins could load next to the mod. We installed nothing and ran no mod code.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Count.&lt;/strong&gt; We left out Anthropic's 4 example mods and 7 whose code the validator could not read. Claude Code would refuse to load those 7 anyway. That leaves &lt;strong&gt;891 community mods from 283 repos&lt;/strong&gt;. For the rarest and most sensitive cases (approving tool calls, credential files), we also read the code by hand.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Results
&lt;/h2&gt;

&lt;p&gt;A mod can fall into several rows.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What the mod can do&lt;/th&gt;
&lt;th&gt;Mods&lt;/th&gt;
&lt;th&gt;Share&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Hooks &lt;code&gt;tool.call&lt;/code&gt;: sees and can change tool calls&lt;/td&gt;
&lt;td&gt;491&lt;/td&gt;
&lt;td&gt;55%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reads environment variables or settings&lt;/td&gt;
&lt;td&gt;389&lt;/td&gt;
&lt;td&gt;44%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Starts programs&lt;/td&gt;
&lt;td&gt;346&lt;/td&gt;
&lt;td&gt;39%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reads files&lt;/td&gt;
&lt;td&gt;325&lt;/td&gt;
&lt;td&gt;36%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hooks &lt;code&gt;tool.call&lt;/code&gt; with no filter: every tool call passes through it&lt;/td&gt;
&lt;td&gt;324&lt;/td&gt;
&lt;td&gt;36%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hooks &lt;code&gt;prompt.submit&lt;/code&gt;: sees the prompts you send&lt;/td&gt;
&lt;td&gt;225&lt;/td&gt;
&lt;td&gt;25%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Writes files&lt;/td&gt;
&lt;td&gt;209&lt;/td&gt;
&lt;td&gt;23%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sends prompts as you&lt;/td&gt;
&lt;td&gt;118&lt;/td&gt;
&lt;td&gt;13%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Calls a model on your plan or key&lt;/td&gt;
&lt;td&gt;107&lt;/td&gt;
&lt;td&gt;12%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Makes network requests&lt;/td&gt;
&lt;td&gt;98&lt;/td&gt;
&lt;td&gt;11%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reads files or environment variables and also makes network requests&lt;/td&gt;
&lt;td&gt;76&lt;/td&gt;
&lt;td&gt;9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hooks &lt;code&gt;tool.check&lt;/code&gt;: can answer permission questions&lt;/td&gt;
&lt;td&gt;48&lt;/td&gt;
&lt;td&gt;5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hooks &lt;code&gt;session.append&lt;/code&gt;: can rewrite stored conversation rows&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;td&gt;4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reads an environment variable named like a secret (KEY, TOKEN, SECRET)&lt;/td&gt;
&lt;td&gt;34&lt;/td&gt;
&lt;td&gt;4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reads files and names a credential path in its code (&lt;code&gt;.env&lt;/code&gt;, &lt;code&gt;~/.ssh&lt;/code&gt;, &lt;code&gt;~/.aws/credentials&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;17&lt;/td&gt;
&lt;td&gt;2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sets environment variables for the commands Claude runs next&lt;/td&gt;
&lt;td&gt;17&lt;/td&gt;
&lt;td&gt;2%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Returns &lt;code&gt;allow&lt;/code&gt; from &lt;code&gt;tool.check&lt;/code&gt;: approves tool calls for you&lt;/td&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;td&gt;1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;None of the above: draws, counts and keeps its own state&lt;/td&gt;
&lt;td&gt;156&lt;/td&gt;
&lt;td&gt;18%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Quick guess: how many of the 891 only draw on the screen and keep their own notes? 156, fewer than one in five.&lt;/p&gt;

&lt;p&gt;A few examples, without names:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Network.&lt;/strong&gt; A mod that lets you continue a session from your phone follows each turn, talks to its relay service over the network, and submits what you type on the phone as your prompt. That is its whole job, and it is also exactly the &lt;code&gt;$.http.fetch&lt;/code&gt; plus &lt;code&gt;$.prompt.submit&lt;/code&gt; combination worth noticing in a mod that claims to do something smaller.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Credential paths.&lt;/strong&gt; All the hits we read were mods reading their own token. One reaction mod for a Telegram bot reads the bot token from a &lt;code&gt;.env&lt;/code&gt; file in its own folder and keeps it in memory. The same &lt;code&gt;$.fs.read&lt;/code&gt; could read any file you can.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every tool call.&lt;/strong&gt; A remote control mod hooks every tool call, every prompt and every stored conversation row, which it needs to drive the session from elsewhere.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Approving for you.&lt;/strong&gt; We read all 11 mods that return &lt;code&gt;allow&lt;/code&gt;. Each one has a condition:

&lt;ul&gt;
&lt;li&gt;Five approve only their own tools or calls from their own plugin.&lt;/li&gt;
&lt;li&gt;Two approve read-only calls, such as reads inside a skill folder.&lt;/li&gt;
&lt;li&gt;One approves a single fixed search command.&lt;/li&gt;
&lt;li&gt;One approves an edit while the session holds a lock on that file.&lt;/li&gt;
&lt;li&gt;One asks an outside command whether the session holds the resource.&lt;/li&gt;
&lt;li&gt;One asks a model to score the risk and approves below a threshold. That is the widest of the eleven.&lt;/li&gt;
&lt;li&gt;None of them approves arbitrary shell commands.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How this compares with the public catalogue
&lt;/h2&gt;

&lt;p&gt;The awesome-claude-code-mods catalogue at &lt;a href="https://mods.aidojo.si/" rel="noopener noreferrer"&gt;mods.aidojo.si&lt;/a&gt; scans GitHub every day. Its November 1 snapshot lists 8,532 mods. It also uses &lt;code&gt;claude plugin validate&lt;/code&gt;, and it gives each mod one reach level, from drawing only up to network access.&lt;/p&gt;

&lt;p&gt;The two samples differ in three ways:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Scope.&lt;/strong&gt; We only took mods from repos with 5 or more stars and a push in the last 30 days, so our 891 are the more used, more active part of the ecosystem.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Counting.&lt;/strong&gt; We count each kind of call and hook separately and read the approval code by hand, where the catalogue gives one level per mod.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Timing.&lt;/strong&gt; We cloned on November 6.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Where the two overlap, they agree:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;521 of our mods appear in the catalogue, and our network flag matches its data on all 521.&lt;/li&gt;
&lt;li&gt;Its whole-catalogue shares sit close to ours: network 9% (ours 11%), starting programs 38% (39%), writing files 18% (23%), sending prompts 11% (13%).&lt;/li&gt;
&lt;li&gt;161 of our 283 repos are missing from that snapshot, all of them created before it.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What to ignore
&lt;/h2&gt;

&lt;p&gt;Stars. The mods that reach furthest are often the useful ones: a remote control, a database approval panel, a kanban board that reads your queue. Reach tells you what to read, and the description tells you whether the reach makes sense.&lt;/p&gt;

&lt;p&gt;Also ignore the idea that a validator pass means a mod is safe. Validation checks that Claude Code can read what the mod will call. Whether those calls are a good idea is still your decision.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Static only.&lt;/strong&gt; This is a static reading of the code. A listed call may sit in a branch that never runs unless you click something. In the other direction, Claude Code refuses to load a mod whose calls it cannot read, so a mod that loads cannot hide a call from this list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Capabilities, not intent.&lt;/strong&gt; We did not judge whether any mod misuses what it asks for, and we read only the rare cases by hand.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One snapshot.&lt;/strong&gt; The sample is mods from active repos with 5 or more stars, as of November 6, 2026. Mods gain new events and calls nearly every week, so these numbers will move.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Reproduce it
&lt;/h2&gt;

&lt;p&gt;The counts above are in &lt;a href="https://agentcraft.dev/data/mods-review-2026-11.json" rel="noopener noreferrer"&gt;mods-review-2026-11.json&lt;/a&gt;, with repo and mod names left out. To check a mod yourself, clone it and run &lt;code&gt;claude plugin validate&lt;/code&gt; on its folder. The docs' &lt;a href="https://code.claude.com/docs/en/plugins/mods/admin" rel="noopener noreferrer"&gt;Review what a mod can do&lt;/a&gt; section lists which calls to look for.&lt;/p&gt;

&lt;p&gt;Looking for something to install? The &lt;a href="https://agentcraft.dev/en/skills/" rel="noopener noreferrer"&gt;AgentCraft skills directory&lt;/a&gt; lists 5,101 Claude Code skills, and every community skill is reviewed before it's listed.&lt;/p&gt;

&lt;p&gt;The full post with the comparison and the data file is on AgentCraft: &lt;a href="https://agentcraft.dev/en/blog/claude-code-mods-review-2026-11/" rel="noopener noreferrer"&gt;https://agentcraft.dev/en/blog/claude-code-mods-review-2026-11/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Disclosure: I'm Adrian Vale and I help run AgentCraft.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>security</category>
      <category>plugins</category>
    </item>
    <item>
      <title>Three checks before you install a Claude Code skill (from a review of 783 skills)</title>
      <dc:creator>Adrian Vale</dc:creator>
      <pubDate>Thu, 05 Nov 2026 16:40:41 +0000</pubDate>
      <link>https://dev.to/adrian2016/three-checks-before-you-install-a-claude-code-skill-from-a-review-of-783-skills-4d5n</link>
      <guid>https://dev.to/adrian2016/three-checks-before-you-install-a-claude-code-skill-from-a-review-of-783-skills-4d5n</guid>
      <description>&lt;p&gt;Before you install a Claude Code skill, open its SKILL.md and check three things.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Does it contain real steps?&lt;/strong&gt; A one-line stub, a pointer to another file, or unreadable ciphertext means there is nothing to install.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Where does your data go?&lt;/strong&gt; Search the file for URLs, keys and sign-up links. Calling an outside service is fine when you know about it and chose it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Would it work on your machine?&lt;/strong&gt; Hard-coded paths, someone else's Slack channel, instructions to switch off the sandbox, or steps for getting past a site's bot checks are all reasons to skip it.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;These come from a review we ran on the AgentCraft directory in November. We collected every Claude Code skill on GitHub that we had yet to list, 783 in total, and dropped 317 of them. That is about two in five.&lt;/p&gt;

&lt;p&gt;The first check alone would have caught 128 of the 317.&lt;/p&gt;

&lt;h2&gt;
  
  
  Stars told us almost nothing
&lt;/h2&gt;

&lt;p&gt;Quick guess: did repo stars help predict which skills got dropped?&lt;/p&gt;

&lt;p&gt;Barely. Dropped skills came from repos with a median of 1,572 stars. Kept skills came from repos with a median of 1,643. And 220 of the 317 dropped skills live in repos with more than 1,000 stars. Popular repos often ship demos, templates and test fixtures next to their real skills, and those were a large share of what got dropped.&lt;/p&gt;

&lt;h2&gt;
  
  
  How the 783 were collected and reviewed
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find repos.&lt;/strong&gt; Repos linked from five large awesome lists of Claude skills, plus eleven GitHub searches on topics such as agent-skills and claude-skills: 1,888 repos.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Filter.&lt;/strong&gt; At least 100 stars, no forks, no archived repos, pushed to in the last year, and no repos that only mirror other people's skills: 1,569 repos, of which 1,414 contain at least one SKILL.md.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pick.&lt;/strong&gt; A few skills per repo depending on stars (20 for 5,000+ stars, down to 2 for 100+), keeping the most starred copy when the same skill name shows up in several repos: 5,169 skills. 783 of them were new to the directory and came back from review in a usable form.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review.&lt;/strong&gt; Claude Haiku read each SKILL.md, eight per call, and either filed it under one of 38 categories or dropped it with a short reason.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The review prompt drops a skill when it exfiltrates data or secrets, harvests credentials, evades bot detection, anti-fraud or platform rules, is malware or offensive hacking against others, is mainly an ad or paid-API funnel, is a placeholder, template, test fixture or near-empty file, is a personal config useful only to its author, or is sexual content.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the 317 were dropped
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Reason&lt;/th&gt;
&lt;th&gt;Skills&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Placeholder, template, demo or test fixture&lt;/td&gt;
&lt;td&gt;108&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Paid API or ad funnel&lt;/td&gt;
&lt;td&gt;65&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Only useful to its author or one project&lt;/td&gt;
&lt;td&gt;46&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Unsafe&lt;/td&gt;
&lt;td&gt;41&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Evades detection or platform rules&lt;/td&gt;
&lt;td&gt;35&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Other (thin aliases, failed auto-extraction)&lt;/td&gt;
&lt;td&gt;14&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Encrypted, so nobody can read it&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Sexual content&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A few examples:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Demo.&lt;/strong&gt; In a repo with about 68,000 stars, one skill only runs the &lt;code&gt;date&lt;/code&gt; command and another fetches the temperature of one fixed city. Great for learning how skills work, useless to install.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Paid API funnel.&lt;/strong&gt; Seven near-identical skills each send you to sign up for the same paid transcript API and paste in a key.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unsafe.&lt;/strong&gt; One skill sends the run's context to an outside service on its own. Nine of the 41 come from two security scanners, NVIDIA SkillSpector and Cisco's skill-scanner, which keep deliberately malicious skills as test cases, including a recipe assistant whose guardrail tells the agent to add cyanide. Those are test samples doing their job. Counting their harmless fixtures too, 25 of the 317 dropped skills come from these two scanners' test folders.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Encrypted.&lt;/strong&gt; Six skills from one repo are stored with git-crypt, so the SKILL.md you would install is ciphertext.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We kept 466. One was a duplicate, so 465 went into the directory.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits worth knowing
&lt;/h2&gt;

&lt;p&gt;Each skill got one pass from one small model, reading only the first 1,800 characters of its SKILL.md.&lt;/p&gt;

&lt;p&gt;We also got an accidental stability test. A name-matching bug sent 266 skills that were already listed through review a second time. Same prompt, same model: 257 kept, 9 dropped. Those 266 are left out of every number above.&lt;/p&gt;

&lt;p&gt;A dropped skill also says nothing bad about its repo. Anthropic's own skills repo has a template skill that was dropped here simply because it gives an agent nothing to do.&lt;/p&gt;

&lt;p&gt;The full write-up, the exact review prompt and a JSON file with all 317 reasons (names removed) are on the original post:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://agentcraft.dev/en/blog/claude-code-skills-review-2026-11/" rel="noopener noreferrer"&gt;https://agentcraft.dev/en/blog/claude-code-skills-review-2026-11/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The 465 kept skills are in the AgentCraft skills directory (5,101 skills, with install commands):&lt;/p&gt;

&lt;p&gt;&lt;a href="https://agentcraft.dev/en/skills/" rel="noopener noreferrer"&gt;https://agentcraft.dev/en/skills/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Disclosure: I'm Adrian Vale and I help run AgentCraft.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>agents</category>
      <category>security</category>
    </item>
    <item>
      <title>An Open AI Prompt Library: 2,600 Image and Video Prompts You Can Download as a Dataset</title>
      <dc:creator>Adrian Vale</dc:creator>
      <pubDate>Fri, 30 Oct 2026 14:26:17 +0000</pubDate>
      <link>https://dev.to/adrian2016/an-open-ai-prompt-library-2600-image-and-video-prompts-you-can-download-as-a-dataset-1f3p</link>
      <guid>https://dev.to/adrian2016/an-open-ai-prompt-library-2600-image-and-video-prompts-you-can-download-as-a-dataset-1f3p</guid>
      <description>&lt;p&gt;Here is my advice for anyone writing prompts for image or video models: start from a prompt that already worked.&lt;/p&gt;

&lt;p&gt;The reason is simple. A good prompt carries a structure, and the structure is the part that is hard to invent. Once you have it, you only swap in your own subject.&lt;/p&gt;

&lt;p&gt;So where do you find prompts that worked? AgentCraft is a free library built for exactly this job.&lt;/p&gt;

&lt;h2&gt;
  
  
  A quick guess
&lt;/h2&gt;

&lt;p&gt;Let me start with a question. A 15 second film clip, 1940s war documentary look, made with MiniMax H3. How long do you think the prompt was?&lt;/p&gt;

&lt;p&gt;One sentence? A short paragraph?&lt;/p&gt;

&lt;p&gt;The answer: the creator split the 15 seconds into timed beats. An establishing shot, rising tension, the firefight, a human moment, then a pull back. Each beat has its own camera move. Below that sit separate blocks for the environment, the characters, the visual style and the physics, plus a long negative prompt.&lt;/p&gt;

&lt;p&gt;That structure is why the clip holds its 35mm look from start to end. You can copy it for any period piece you want to make.&lt;/p&gt;

&lt;p&gt;This prompt came from @RuzainaMeer on X. AgentCraft keeps it word for word, next to the clip it produced, with a link back to the post.&lt;/p&gt;

&lt;h2&gt;
  
  
  What AgentCraft has
&lt;/h2&gt;

&lt;p&gt;The whole library is open, and the numbers are large.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Video prompts, 940.&lt;/strong&gt; Seedance 2.5, MiniMax H3, Wan 3.0, Kling 4.0, Veo and Gemini Omni, grouped by type: cinematic, action, sci-fi, anime, ads and more.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Image prompts, 1,307.&lt;/strong&gt; GPT Image 2 and 2.5 and Nano Banana, grouped by use: posters, portraits, infographics, product shots, UI mockups, logos.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Agent prompts, 353.&lt;/strong&gt; System and role prompts for Claude Code, Codex and other coding agents, grouped by field.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code skills, 4,636.&lt;/strong&gt; Each one read and screened before it was listed, with its install command.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Every video and image prompt shows the output it made and the place it came from, either the model maker's guide or a creator's post. You see the result first, then decide whether the prompt is worth your time.&lt;/p&gt;

&lt;p&gt;Links: &lt;a href="https://agentcraft.dev/en/prompts/video/" rel="noopener noreferrer"&gt;video prompts&lt;/a&gt;, &lt;a href="https://agentcraft.dev/en/prompts/image/" rel="noopener noreferrer"&gt;image prompts&lt;/a&gt;, &lt;a href="https://agentcraft.dev/en/prompts/agents/" rel="noopener noreferrer"&gt;agent prompts&lt;/a&gt;, &lt;a href="https://agentcraft.dev/en/skills/" rel="noopener noreferrer"&gt;skills&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The open data
&lt;/h2&gt;

&lt;p&gt;This is the part I like most. Everything behind the site is public, so you can take the data and build on it.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;&lt;a href="https://github.com/ai-boost/awesome-image-video-prompts" rel="noopener noreferrer"&gt;awesome-image-video-prompts&lt;/a&gt;&lt;/strong&gt; on GitHub, 3.4k stars. The video and image prompts, with outputs and sources.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;a href="https://github.com/ai-boost/awesome-prompts" rel="noopener noreferrer"&gt;awesome-prompts&lt;/a&gt;&lt;/strong&gt; on GitHub, 9.2k stars. Prompts, frameworks and papers for LLMs and agents.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;a href="https://huggingface.co/datasets/Adrian2016/awesome-prompts" rel="noopener noreferrer"&gt;awesome-prompts on Hugging Face&lt;/a&gt;&lt;/strong&gt;, 3,149 prompts in two subsets. The engineering subset has 372 system prompts; 73 of them cite arXiv papers by id. The library subset has 2,777 everyday prompts for writing, marketing, study and design, each in English and Chinese.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Loading the dataset takes two lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;datasets&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;load_dataset&lt;/span&gt;

&lt;span class="n"&gt;eng&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;load_dataset&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Adrian2016/awesome-prompts&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;engineering&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;split&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;train&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;lib&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;load_dataset&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Adrian2016/awesome-prompts&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;library&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;split&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;train&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A prompt dataset with sources attached is handy for evals, for fine-tuning a prompt writer, or for studying which patterns models follow.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three patterns worth stealing
&lt;/h2&gt;

&lt;p&gt;Browse the library for a while and a few patterns keep showing up. Here are three that are easy to reuse!&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Give video a timeline.&lt;/strong&gt; Write the clip as timed beats with a camera move for each, as in the war clip above. The model then plans the shots in order.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Give images a count.&lt;/strong&gt; One GPT Image 2 prompt asks for a town scene with exactly 100 elements, then asks the model to list every item. The model plans the picture as an inventory, and you get a dense illustrated town under a numbered list.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Name the layout.&lt;/strong&gt; Another prompt asks for a catalog of 1990s anime character styles: 16 labeled archetypes in a grid, one per panel, at 16:9. One sentence, and the result is a clean reference sheet.&lt;/p&gt;

&lt;p&gt;The common thread: tell the model how to organize the work, and it fills in the rest well.&lt;/p&gt;

&lt;h2&gt;
  
  
  Lesson
&lt;/h2&gt;

&lt;p&gt;Good prompts are templates. Borrow the structure, keep your own subject.&lt;/p&gt;

&lt;p&gt;Which model are you prompting most these days, and what structure works for you?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>promptengineering</category>
      <category>opensource</category>
      <category>machinelearning</category>
    </item>
  </channel>
</rss>
