<?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: Nirali</title>
    <description>The latest articles on DEV Community by Nirali (@nirali_toolbit).</description>
    <link>https://dev.to/nirali_toolbit</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%2F4203648%2F65c48142-2f37-4c51-b179-a57c1965e20b.png</url>
      <title>DEV Community: Nirali</title>
      <link>https://dev.to/nirali_toolbit</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/nirali_toolbit"/>
    <language>en</language>
    <item>
      <title>Over 40% of Pull Requests Now Start With an Agent. Review Them Like It.</title>
      <dc:creator>Nirali</dc:creator>
      <pubDate>Tue, 03 Nov 2026 04:35:00 +0000</pubDate>
      <link>https://dev.to/toolbit_ai/over-40-of-pull-requests-now-start-with-an-agent-review-them-like-it-3d44</link>
      <guid>https://dev.to/toolbit_ai/over-40-of-pull-requests-now-start-with-an-agent-review-them-like-it-3d44</guid>
      <description>&lt;p&gt;GitHub's new Octoverse numbers came out on 28 October, and one line stuck with me: "As of September, developers start more than 40% of pull requests with an agent." In March that number was 17%.&lt;/p&gt;

&lt;p&gt;So if you review code for a team, a big chunk of what lands in your queue probably wasn't typed by the person whose name is on it. That doesn't make it bad code. GitHub says 90% of agent PRs were merged within two weeks in September, against about 86% for PRs with no agent involved. But it does change what's worth checking.&lt;/p&gt;

&lt;p&gt;Here's what GitHub already does for you, what it just changed, and the checks I'd add on top.&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%2Fkxlicsw2n69gp6eotxlh.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%2Fkxlicsw2n69gp6eotxlh.png" alt="Bar chart: PRs started with an agent rose from 17% in March to over 40% in September; agent PRs merged within two weeks 90% vs about 86%; 87% of agent PRs change more than one file vs 79%" width="800" height="497"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Source: GitHub Octoverse 2026.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  What GitHub's coding agent already protects
&lt;/h2&gt;

&lt;p&gt;If you use Copilot's cloud agent, a few guardrails are on by default. Here's what GitHub's docs say it handles, and what's still on you:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What GitHub does&lt;/th&gt;
&lt;th&gt;Default&lt;/th&gt;
&lt;th&gt;What's still on you&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Only people with write access can start the agent&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;Who has write access in the first place&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Agent pushes to a single branch (normally &lt;code&gt;copilot/...&lt;/code&gt;) and your branch protections still apply&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;That your protections and required checks are actually set&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The person who asked for the PR &lt;strong&gt;can't approve it&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;A real second review, not a quick thumbs-up&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Actions workflows don't run on Copilot's pushes until a maintainer clicks &lt;strong&gt;Approve and run workflows&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;Reading any &lt;code&gt;.github/workflows/&lt;/code&gt; change before you click&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Secret scanning, CodeQL and an advisory-database dependency check run on its code&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;Whether new packages are the ones you meant&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Copilot approvals&lt;/strong&gt; count toward required reviews&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;Off&lt;/strong&gt; (new, path-scoped)&lt;/td&gt;
&lt;td&gt;Which paths, if any, you let it approve&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;That's a decent baseline. The second-reviewer rule alone stops the "I asked for it, I approved it, it's merged" loop.&lt;/p&gt;
&lt;h2&gt;
  
  
  The new setting to decide on this week
&lt;/h2&gt;

&lt;p&gt;On the same day, GitHub made &lt;strong&gt;Copilot approvals&lt;/strong&gt; generally available. When it's on, Copilot can approve a PR and "that approval counts toward your repository's required approvals." It's off by default, and admins pick which file paths it's allowed to approve.&lt;/p&gt;

&lt;p&gt;So now you have an odd mix: the person who asked the agent for a PR can't approve it, but Copilot's own approval can count on the paths you allow. That might be fine for docs or test fixtures. I wouldn't let it near auth, billing, migrations or &lt;code&gt;.github/workflows/&lt;/code&gt; without a person looking too. Whatever you pick, write it down so the team knows which paths a bot can sign off.&lt;/p&gt;
&lt;h2&gt;
  
  
  Five things I'd check on every agent PR
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. Did the tests change along with the code?&lt;/strong&gt;&lt;br&gt;
Researchers behind ImpossibleBench found agents "may delete failing tests rather than fix the underlying bug." So when a fix and its tests change in the same PR, read the test diff first:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git diff &lt;span class="nt"&gt;--stat&lt;/span&gt; origin/main...HEAD &lt;span class="nt"&gt;--&lt;/span&gt; &lt;span class="s1"&gt;'*test*'&lt;/span&gt; &lt;span class="s1"&gt;'*spec*'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;(It's a rough filter, so it'll also catch names like &lt;code&gt;latest.js&lt;/code&gt;. Good enough for a first look.)&lt;/p&gt;

&lt;p&gt;If an assertion got looser or a test disappeared, ask why before you look at anything else.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Any new dependencies?&lt;/strong&gt;&lt;br&gt;
A USENIX Security 2025 paper found an average hallucinated-package rate of at least 5.2% for commercial models and 21.7% for open-source ones. Attackers can register those names. Check every new line in your manifest actually points at the package you think it does:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git diff origin/main...HEAD &lt;span class="nt"&gt;--&lt;/span&gt; &lt;span class="s1"&gt;'*package.json'&lt;/span&gt; &lt;span class="s1"&gt;'*requirements*.txt'&lt;/span&gt; &lt;span class="s1"&gt;'*pyproject.toml'&lt;/span&gt; &lt;span class="s1"&gt;'*go.mod'&lt;/span&gt; &lt;span class="s1"&gt;'*Cargo.toml'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;3. Anything under &lt;code&gt;.github/workflows/&lt;/code&gt;?&lt;/strong&gt;&lt;br&gt;
GitHub's own review guide calls this out. A workflow change runs with your repo's permissions, so it deserves a slower read than the feature code.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Where did the task come from?&lt;/strong&gt;&lt;br&gt;
In 2025 Invariant Labs showed an agent could be hijacked "via a malicious GitHub Issue" and end up leaking private data in a PR it opened itself. If the agent worked from an issue that an outsider wrote, read the issue too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Is it doing more than it was asked?&lt;/strong&gt;&lt;br&gt;
GitHub says about 87% of agent-started PRs touch more than one file, compared with 79% for PRs with no agent detected. Wide PRs hide small surprises. If the task was "fix the date parser" and the diff also renames a config key, split it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not just trust it?
&lt;/h2&gt;

&lt;p&gt;Because the research still says "check." Veracode tested over 100 models in 2025 and 45% of the code samples failed security tests. Google's 2025 DORA report found 30% of developers have little or no trust in AI-generated code, and that AI use still has "a negative relationship with software delivery stability." None of that means agents are useless. It means review is where the value gets decided, which is pretty much what GitHub's own Octoverse post says: "Review is where developers shape what comes back."&lt;/p&gt;

&lt;p&gt;If you're looking at review tools to help with the load, Toolbit has a &lt;a href="https://www.toolbit.ai/category/code-review" rel="noopener noreferrer"&gt;page of AI code review tools&lt;/a&gt; you can compare.&lt;/p&gt;

&lt;p&gt;What's on your team's checklist for agent PRs that isn't on mine?&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This article was created with the help of AI and reviewed by the author for accuracy. Disclosure: I work at Toolbit, linked above.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Sources&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;GitHub Blog, Octoverse 2026 (28 Oct 2026): &lt;a href="https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/" rel="noopener noreferrer"&gt;https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;GitHub Changelog, Copilot code review Max and approvals (28 Oct 2026): &lt;a href="https://github.blog/changelog/2026-10-28-copilot-code-review-merge-with-confidence-with-max-and-approvals" rel="noopener noreferrer"&gt;https://github.blog/changelog/2026-10-28-copilot-code-review-merge-with-confidence-with-max-and-approvals&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;GitHub Docs, Copilot cloud agent risks and mitigations: &lt;a href="https://docs.github.com/en/copilot/concepts/agents/cloud-agent/risks-and-mitigations" rel="noopener noreferrer"&gt;https://docs.github.com/en/copilot/concepts/agents/cloud-agent/risks-and-mitigations&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;GitHub Docs, Reviewing Copilot pull requests: &lt;a href="https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent/review-copilot-prs" rel="noopener noreferrer"&gt;https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent/review-copilot-prs&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;ImpossibleBench, arXiv 2510.20270: &lt;a href="https://arxiv.org/abs/2510.20270" rel="noopener noreferrer"&gt;https://arxiv.org/abs/2510.20270&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Spracklen et al., package hallucinations, USENIX Security 2025: &lt;a href="https://arxiv.org/abs/2406.10279" rel="noopener noreferrer"&gt;https://arxiv.org/abs/2406.10279&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Invariant Labs, GitHub MCP exploit (26 May 2025): &lt;a href="https://invariantlabs.ai/blog/mcp-github-vulnerability" rel="noopener noreferrer"&gt;https://invariantlabs.ai/blog/mcp-github-vulnerability&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Veracode, 2025 GenAI Code Security Report: &lt;a href="https://www.veracode.com/blog/genai-code-security-report/" rel="noopener noreferrer"&gt;https://www.veracode.com/blog/genai-code-security-report/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Google Cloud, 2025 DORA report: &lt;a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report" rel="noopener noreferrer"&gt;https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>github</category>
      <category>codereview</category>
      <category>security</category>
    </item>
    <item>
      <title>Prompt Caching Cost Is What Decides Your AI Coding Agent Bill</title>
      <dc:creator>Nirali</dc:creator>
      <pubDate>Mon, 26 Oct 2026 09:59:08 +0000</pubDate>
      <link>https://dev.to/toolbit_ai/prompt-caching-cost-is-what-decides-your-ai-coding-agent-bill-50ad</link>
      <guid>https://dev.to/toolbit_ai/prompt-caching-cost-is-what-decides-your-ai-coding-agent-bill-50ad</guid>
      <description>&lt;p&gt;A coding agent sends your whole conversation back to the model on every single turn. So the price that decides your bill is often the smallest number in the pricing table: the prompt caching cost.&lt;/p&gt;

&lt;p&gt;Max Woolf's post &lt;a href="https://minimaxir.com/2026/10/cheap-llms/" rel="noopener noreferrer"&gt;OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?&lt;/a&gt; is getting a lot of discussion on Hacker News right now. It walks through a summer of price cuts. One line stuck with me. He writes that when you use agents for coding, cache read costs are "the vast majority (96% - 99%) of the total token expenditure." I wanted to see what that means for a real AI coding agent cost, using today's pricing pages instead of launch headlines.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why agents live on cache reads
&lt;/h2&gt;

&lt;p&gt;Every time an agent calls the model, it sends the system prompt, the tool definitions, every file it has read and every message so far. Then it adds a little new content at the end: a tool result, or your next message.&lt;/p&gt;

&lt;p&gt;Providers cache that repeated prefix. Reading it back from the cache is much cheaper than sending it fresh:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;On Claude Opus 5.5, a cache hit costs 5% of the base input price.&lt;/li&gt;
&lt;li&gt;On GPT-6.1 Sol, it's also 5% ($0.10 vs $2.00).&lt;/li&gt;
&lt;li&gt;On Claude Fable 5.1, it's 2.5%.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The new tokens at the end usually get &lt;em&gt;written&lt;/em&gt; to the cache. That costs a bit more than normal input: 1.25x on both Anthropic's default 5-minute cache and OpenAI's default 30-minute cache.&lt;/p&gt;

&lt;h2&gt;
  
  
  LLM API pricing 2026: the numbers that matter
&lt;/h2&gt;

&lt;p&gt;Here is what the official pricing pages say today. Prices are USD per million tokens, standard tier.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Input&lt;/th&gt;
&lt;th&gt;Cache read&lt;/th&gt;
&lt;th&gt;Cache write (default)&lt;/th&gt;
&lt;th&gt;Output&lt;/th&gt;
&lt;th&gt;Note&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Claude Opus 5&lt;/td&gt;
&lt;td&gt;$5.00&lt;/td&gt;
&lt;td&gt;$0.50&lt;/td&gt;
&lt;td&gt;$6.25 (5 min)&lt;/td&gt;
&lt;td&gt;$25&lt;/td&gt;
&lt;td&gt;previous Opus&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Opus 5.5&lt;/td&gt;
&lt;td&gt;$4.00&lt;/td&gt;
&lt;td&gt;$0.20&lt;/td&gt;
&lt;td&gt;$5.00 (5 min)&lt;/td&gt;
&lt;td&gt;$20&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Fable 5.1&lt;/td&gt;
&lt;td&gt;$10.00&lt;/td&gt;
&lt;td&gt;$0.25&lt;/td&gt;
&lt;td&gt;$12.50 (5 min)&lt;/td&gt;
&lt;td&gt;$50&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-5.6 Sol&lt;/td&gt;
&lt;td&gt;$4.00&lt;/td&gt;
&lt;td&gt;$0.40&lt;/td&gt;
&lt;td&gt;$5.00 (30 min)&lt;/td&gt;
&lt;td&gt;$20&lt;/td&gt;
&lt;td&gt;promotional price&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-6.1 Sol&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;$0.10&lt;/td&gt;
&lt;td&gt;$2.50 (30 min)&lt;/td&gt;
&lt;td&gt;$10&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;$0.10&lt;/td&gt;
&lt;td&gt;not listed (implicit caching only)&lt;/td&gt;
&lt;td&gt;$10&lt;/td&gt;
&lt;td&gt;until 30 Nov 2026&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Checked 26 Oct 2026 on the &lt;a href="https://platform.claude.com/docs/en/about-claude/pricing" rel="noopener noreferrer"&gt;Anthropic&lt;/a&gt;, &lt;a href="https://developers.openai.com/api/docs/pricing" rel="noopener noreferrer"&gt;OpenAI&lt;/a&gt; and &lt;a href="https://ai.google.dev/gemini-api/docs/pricing" rel="noopener noreferrer"&gt;Google&lt;/a&gt; pricing pages. OpenAI prices are for prompts up to 272K input tokens.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;(I work at Toolbit, and we keep a running comparison of these prices &lt;a href="https://www.toolbit.ai/blog/opus-5-5-gpt-6-sol-luna-price-war" rel="noopener noreferrer"&gt;here&lt;/a&gt;.)&lt;/p&gt;

&lt;p&gt;Look at Opus 5 to Opus 5.5. Input and output dropped 20%. Cache reads dropped 60%, from $0.50 to $0.20. If you only read the headline, you'd plan for a 20% saving. An agent gets a lot more than that.&lt;/p&gt;

&lt;h2&gt;
  
  
  A worked example: one hour of agent work
&lt;/h2&gt;

&lt;p&gt;These are made-up but realistic assumptions for one coding session:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;160 model calls in about an hour&lt;/li&gt;
&lt;li&gt;the context averages about 100K tokens over the session (it grows, and the agent compacts it now and then), so each call re-reads about 100K cached tokens on average&lt;/li&gt;
&lt;li&gt;about 2K new tokens per call (tool output, your messages), written to the cache&lt;/li&gt;
&lt;li&gt;about 500 output tokens per call (mostly tool calls and short replies)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That adds up to 16M cache-read tokens, 0.32M cache-write tokens and 0.08M output tokens. Cache reads are 97.6% of all tokens, right inside Woolf's range.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Cache reads&lt;/th&gt;
&lt;th&gt;Cache writes&lt;/th&gt;
&lt;th&gt;Output&lt;/th&gt;
&lt;th&gt;Session total&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Opus 5&lt;/td&gt;
&lt;td&gt;$8.00&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$12.00&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Opus 5.5&lt;/td&gt;
&lt;td&gt;$3.20&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$6.40&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-5.6 Sol (promo)&lt;/td&gt;
&lt;td&gt;$6.40&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$9.60&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-6.1 Sol&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$0.80&lt;/td&gt;
&lt;td&gt;$0.80&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$3.20&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon (until 30 Nov)&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$0.64*&lt;/td&gt;
&lt;td&gt;$0.80&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$3.04&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon (from 1 Dec)&lt;/td&gt;
&lt;td&gt;$3.20&lt;/td&gt;
&lt;td&gt;$1.28*&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$6.08&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Google lists no cache-write price for Argon, so I priced the new tokens at the normal input rate. Its caching is implicit, so your hit rate isn't guaranteed.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Three things stand out.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Opus 5 to 5.5 is a 47% cut for this session, not 20%.&lt;/strong&gt; The 60% cache-read cut does most of the work. A headline-only estimate would have said $9.60. That lines up with Woolf's estimate of a 40-60% total cost cut.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cache reads are the biggest line item, but not 96-99% of the money here.&lt;/strong&gt; They're 97.6% of the tokens. In dollars they're 50-67% of the bill, because writes and output cost far more per token. If your agent writes long answers, output matters more than you'd think.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Caching itself is the real saving.&lt;/strong&gt; Run the same session on Opus 5.5 with no caching and the input alone is 16.32M × $4 = $65.28. That's $66.88 in total, against $6.40 with caching.&lt;/li&gt;
&lt;/ol&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%2Faa1ies67vqduffw9z7u2.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%2Faa1ies67vqduffw9z7u2.png" alt="Stacked bar chart for an example Claude Opus 5.5 coding session: cache reads are 97.6% of tokens but 50% of the $6.40 cost; cache writes and output are 25% each" width="799" height="411"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Which cuts are temporary
&lt;/h2&gt;

&lt;p&gt;Not every price in that table will still be there in January.&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%2Foye1rx0widl5mu7wtkgf.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%2Foye1rx0widl5mu7wtkgf.png" alt="Timeline of promotional API prices ending: GPT-5.6 Sol promo at least through 21 Nov; Gemini 4 Argon $2/$10 until 30 Nov, then $4/$20; Gemini 3.8 Flash $0.75/$3.75 until 31 Dec, then $1.50/$7.50" width="799" height="434"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GPT-5.6 Sol:&lt;/strong&gt; OpenAI calls its current price promotional, "available at least through November 21, 2026."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemini 4 Argon:&lt;/strong&gt; $2 input / $10 output through 30 Nov 2026, then $4 / $20 from 1 Dec. Cache reads go from $0.10 to $0.20. The example session doubles.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemini 3.8 Flash:&lt;/strong&gt; $0.75 / $3.75 through 31 Dec 2026, then $1.50 / $7.50 from 1 Jan 2027.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Sonnet 5&lt;/strong&gt; went the other way. Its $2 / $10 launch price was due to rise to $3 / $15 on 1 Sep 2026. Anthropic made $2 / $10 the standard price instead.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you pick a model on price, put those dates in your calendar.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to estimate your own prompt caching cost
&lt;/h2&gt;

&lt;p&gt;You don't need my assumptions. Your usage logs have the real numbers.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Pull token counts per call.&lt;/strong&gt; Anthropic's API returns &lt;code&gt;cache_read_input_tokens&lt;/code&gt; and &lt;code&gt;cache_creation_input_tokens&lt;/code&gt; in the usage block. OpenAI reports &lt;code&gt;cached_tokens&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Split them into four buckets:&lt;/strong&gt; cache reads, cache writes, uncached input and output.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multiply each bucket by its own price.&lt;/strong&gt; Don't use the input price for everything.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check your hit rate.&lt;/strong&gt; Woolf says the cache tends to be hit around 90-100% of the time, depending on the provider. If yours is much lower, fix that before you switch models.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Watch for cache expiry.&lt;/strong&gt; Anthropic's default cache lives for 5 minutes and is refreshed each time it's read. Step away for a coffee and the next call writes the whole context again. On Opus 5.5, rewriting 100K tokens costs $0.50, while reading them costs $0.02. OpenAI's default is 30 minutes. Both offer longer lifetimes for a higher write price: 1 hour at 2x input on Anthropic, 12 hours at 2.5x on OpenAI.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Then redo the math for the date each temporary price ends.&lt;/p&gt;

&lt;h2&gt;
  
  
  The short version
&lt;/h2&gt;

&lt;p&gt;Input and output prices make the news. For coding agents, the cache-read column and your hit rate decide most of the bill. Cheaper cache reads are why Opus 5.5 costs about half as much as Opus 5 in the example above, even though its headline prices only fell 20%.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI tools helped me research and draft this post. I checked every number against the linked sources on 26 Oct 2026.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;What does your agent's split look like? If you've checked your usage logs, I'd love to know how much of your bill is cache reads and how much is output.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>agents</category>
      <category>discuss</category>
    </item>
    <item>
      <title>Open-Weight AI Models You Can Run Yourself: October 2026 Edition</title>
      <dc:creator>Nirali</dc:creator>
      <pubDate>Mon, 26 Oct 2026 01:54:09 +0000</pubDate>
      <link>https://dev.to/nirali_toolbit/open-weight-ai-models-you-can-run-yourself-october-2026-edition-3ndo</link>
      <guid>https://dev.to/nirali_toolbit/open-weight-ai-models-you-can-run-yourself-october-2026-edition-3ndo</guid>
      <description>&lt;p&gt;Open-weight AI models are easy to download and much harder to fit in memory. Meta's Muse Spark Open, released on 15 October 2026, shows the gap clearly: 816B parameters, a 1M-token context window, and, &lt;a href="https://www.engadget.com/2287833/meta-built-a-huge-open-weight-ai-model-for-macs-but-youll-need-1tb-of-ram-to-run-it/" rel="noopener noreferrer"&gt;per Engadget&lt;/a&gt;, about 1 TB of RAM to run it locally.&lt;/p&gt;

&lt;p&gt;Below: what "open-weight" means, this month's releases side by side, a Python memory estimator, license checks, and hardware sizing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Open-weight vs open source: what you actually get
&lt;/h2&gt;

&lt;p&gt;An open-weight model gives you the trained parameters (the "weights"). You can run it on your own hardware and, if the license allows, fine-tune it.&lt;/p&gt;

&lt;p&gt;That is narrower than open source AI. The Open Source Initiative's &lt;a href="https://opensource.org/ai/open-source-ai-definition" rel="noopener noreferrer"&gt;Open Source AI Definition 1.0&lt;/a&gt; also asks for detailed information about the training data and the complete code used to train and run the system. OSI's &lt;a href="https://opensource.org/ai/open-weights" rel="noopener noreferrer"&gt;explainer on open weights&lt;/a&gt; says they "differ significantly from Open Source AI" because that code and data detail is missing.&lt;/p&gt;

&lt;p&gt;In short: you can download and run the model. The license decides everything else.&lt;/p&gt;

&lt;h2&gt;
  
  
  October 2026 open-weight models at a glance
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Released&lt;/th&gt;
&lt;th&gt;Total parameters&lt;/th&gt;
&lt;th&gt;Active per token&lt;/th&gt;
&lt;th&gt;Context&lt;/th&gt;
&lt;th&gt;License&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Muse Spark Open (Meta)&lt;/td&gt;
&lt;td&gt;15 Oct 2026&lt;/td&gt;
&lt;td&gt;816B (MoE)&lt;/td&gt;
&lt;td&gt;42B&lt;/td&gt;
&lt;td&gt;1M tokens&lt;/td&gt;
&lt;td&gt;Muse Model License (custom)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DeepSeek V4.1 Flash 1017&lt;/td&gt;
&lt;td&gt;17 Oct 2026&lt;/td&gt;
&lt;td&gt;552B backbone (MoE) + 196B Engram memory&lt;/td&gt;
&lt;td&gt;8B prefill, 16B decode&lt;/td&gt;
&lt;td&gt;Up to 1M tokens&lt;/td&gt;
&lt;td&gt;MIT&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Granite Switch 4.1 8B (IBM)&lt;/td&gt;
&lt;td&gt;14 Oct 2026&lt;/td&gt;
&lt;td&gt;8B base model&lt;/td&gt;
&lt;td&gt;Not listed&lt;/td&gt;
&lt;td&gt;128K (base model)&lt;/td&gt;
&lt;td&gt;Apache 2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Qwen3.8-27B (for comparison)&lt;/td&gt;
&lt;td&gt;August 2026&lt;/td&gt;
&lt;td&gt;27B (dense)&lt;/td&gt;
&lt;td&gt;All 27B&lt;/td&gt;
&lt;td&gt;262,144 native, up to 1M&lt;/td&gt;
&lt;td&gt;Apache 2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;All four have downloadable weights on Hugging Face; Muse Spark Open's require an access request.&lt;/p&gt;

&lt;h3&gt;
  
  
  Muse Spark Open: a desktop model, if your desktop has 1 TB
&lt;/h3&gt;

&lt;p&gt;Muse Spark Open is a mixture-of-experts (MoE) model, which routes each token through a few "expert" sub-networks. Its &lt;a href="https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B" rel="noopener noreferrer"&gt;model card&lt;/a&gt; lists 256 experts, 8 active per token.&lt;/p&gt;

&lt;p&gt;Meta's &lt;a href="https://research.meta.ai/blog/muse-spark-open" rel="noopener noreferrer"&gt;launch post&lt;/a&gt; says the model was optimized for two Apple M5 Ultra Mac Studios running MLX. Each machine's GPU can use up to 512 GB of unified memory, so the pair adds up to a terabyte.&lt;/p&gt;

&lt;p&gt;On GPUs, the model card says you need at least 700GB of VRAM for the NVFP4 version and 2TB for BF16. Downloading requires an access request with your full legal name, date of birth and organization. For another reference, see &lt;a href="https://www.toolbit.ai/updates/models/muse-spark-open" rel="noopener noreferrer"&gt;Muse Spark Open on Toolbit&lt;/a&gt;. (I work on Toolbit.)&lt;/p&gt;

&lt;h3&gt;
  
  
  DeepSeek V4.1 Flash 1017: huge total, small active
&lt;/h3&gt;

&lt;p&gt;DeepSeek &lt;a href="https://api-docs.deepseek.com/news/news261017" rel="noopener noreferrer"&gt;released V4.1 Flash 1017&lt;/a&gt; on 17 October, with weights on Hugging Face and ModelScope. Its &lt;a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017" rel="noopener noreferrer"&gt;model card&lt;/a&gt; calls it a minor update to V4.1 Flash and puts it under the MIT license.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" rel="noopener noreferrer"&gt;V4.1 Flash model card&lt;/a&gt; lists 552B backbone parameters, with 8B active per token during prefill (reading your prompt) and 16B during decode (generating output). DeepSeek says its design cuts the global KV cache to 890 bytes per token, under 1 GB for 1M tokens.&lt;/p&gt;

&lt;h3&gt;
  
  
  Smaller options: Granite Switch 4.1 and Qwen3.8-27B
&lt;/h3&gt;

&lt;p&gt;IBM's &lt;a href="https://huggingface.co/ibm-granite/granite-switch-4.1-8b" rel="noopener noreferrer"&gt;Granite Switch 4.1 8B&lt;/a&gt; lists a release date of 14 October. It's built on the Granite 4.1 8B base model (8B parameters, 128K context), uses Apache 2.0, and also comes in a 30B version. IBM also publishes &lt;a href="https://huggingface.co/ibm-granite/granite-switch-4.1-8b-GGUF" rel="noopener noreferrer"&gt;GGUF files&lt;/a&gt; at several quantization levels.&lt;/p&gt;

&lt;p&gt;For mid-size, &lt;a href="https://huggingface.co/Qwen/Qwen3.8-27B" rel="noopener noreferrer"&gt;Qwen3.8-27B&lt;/a&gt; (August) is a 27B dense model with 262,144 tokens of native context, extensible to 1M, under Apache 2.0.&lt;/p&gt;

&lt;h2&gt;
  
  
  How much memory do open-weight AI models need?
&lt;/h2&gt;

&lt;p&gt;Hugging Face's &lt;a href="https://huggingface.co/docs/transformers/llm_tutorial_optimization" rel="noopener noreferrer"&gt;guide to optimizing LLMs&lt;/a&gt; gives a simple rule: a model with X billion parameters needs roughly 2 × X GB of VRAM in bfloat16, or 4 × X GB in float32. Generalized:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;weights (GB) ≈ parameters (billions) × bits per weight ÷ 8&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That's the floor. Two things sit on top of it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;KV cache.&lt;/strong&gt; Stored keys and values for every token in context. The guide counts them as 2 × sequence length × heads × head dimension × layers, so the cache grows linearly with context. Grouped-query attention (GQA) uses fewer key-value heads, which shrinks it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Runtime overhead.&lt;/strong&gt; Activations, buffers and framework memory.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Quantization lowers the bits per weight. The guide says 8-bit or 4-bit weights work without a significant loss in performance, but warns that quantization trades memory against accuracy and sometimes speed. Real formats add a little: &lt;a href="https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/" rel="noopener noreferrer"&gt;NVIDIA's NVFP4&lt;/a&gt; stores one FP8 scale per 16 values, which works out to about 4.5 bits per weight.&lt;/p&gt;

&lt;p&gt;The guide's own test shows the gap: it expects about 31 GB for OctoCoder in bfloat16, so 4-bit should be about 7.75 GB, yet it measured "just a bit over 9GB."&lt;/p&gt;

&lt;h3&gt;
  
  
  A quick Python memory estimator
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;weights_gb&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;params_billion&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;bits_per_weight&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;Weights only, in GB (1 GB = 1e9 bytes).&lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;params_billion&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;bits_per_weight&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;


&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;kv_cache_gb&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tokens&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;layers&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;kv_heads&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;head_dim&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;bytes_per_value&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;Keys + values for every token, layer and KV head (2 bytes = BF16).&lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;tokens&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;layers&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;kv_heads&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;head_dim&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;bytes_per_value&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mf"&gt;1e9&lt;/span&gt;


&lt;span class="n"&gt;MARGIN&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;1.2&lt;/span&gt;  &lt;span class="c1"&gt;# +20% for cache, activations and runtime (an assumption)
&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;label&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;params&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;[(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;8B&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&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;27B&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;27&lt;/span&gt;&lt;span class="p"&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;816B&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;816&lt;/span&gt;&lt;span class="p"&gt;)]:&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;bits&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;16&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;4&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;weights_gb&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;params&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;bits&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;label&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;&lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; @ &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;bits&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;-bit: ~&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;w&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="p"&gt;,.&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; GB weights, ~&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;MARGIN&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="p"&gt;,.&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; GB with margin&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="c1"&gt;# Muse Spark Open model card: 62 layers, 16 KV heads, head dim 64
&lt;/span&gt;&lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Full 1M-token BF16 cache: ~&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="nf"&gt;kv_cache_gb&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1_000_000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;62&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;16&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;64&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="p"&gt;,.&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; GB&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 27B model needs about 13.5 GB of weights at 4-bit, and 816B at BF16 lands near 1,958 GB with the margin, close to the 2TB on Meta's model card.&lt;/p&gt;

&lt;p&gt;The last line uses Muse Spark Open's card values. If every layer caches every token, a full 1M-token context adds roughly 254 GB on top of about 408 GB of 4-bit weights.&lt;/p&gt;

&lt;p&gt;For MoE models, always use &lt;strong&gt;total&lt;/strong&gt; parameters, not active ones.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check in a license before commercial use
&lt;/h2&gt;

&lt;p&gt;Not legal advice, but check these first:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Is it a standard license?&lt;/strong&gt; DeepSeek's MIT license lets you use, modify and sell copies as long as you keep the copyright and license notice. Apache 2.0 (Qwen3.8-27B, Granite Switch) includes an explicit patent license. Custom licenses need a closer read.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Are there attribution rules?&lt;/strong&gt; The Muse Model License requires products using the model to prominently display "Built with Muse Spark Open."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Is there a scale or revenue trigger?&lt;/strong&gt; If you offer model access to third parties via API and your company's gross revenue tops $20 million in any 12-month period, Meta's license requires notifying Meta and signing a separate agreement. Internal use, and powering features of your own product that isn't a model-as-a-service, are exempt.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What does the usage policy ban?&lt;/strong&gt; Muse's license includes a usage policy with a list of prohibited uses, and says the model isn't intended for anyone under 18.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Do obligations follow your fine-tunes?&lt;/strong&gt; Muse's definition of "Derivatives" covers models fine-tuned or distilled using the model or its outputs.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Meta itself calls the commercial terms limited. Have a lawyer read the full text before you ship.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which size fits your hardware?
&lt;/h2&gt;

&lt;p&gt;Weights-only math, so leave room for context:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Laptop with 16 to 32 GB of RAM or unified memory:&lt;/strong&gt; 8B-class models such as Granite Switch 4.1 8B, at 4-bit (about 4 GB) or 8-bit (about 8 GB). &lt;a href="https://github.com/ggml-org/llama.cpp" rel="noopener noreferrer"&gt;llama.cpp&lt;/a&gt; treats Apple silicon as "a first-class citizen" and also supports x86 CPUs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Single-GPU workstation with 24 to 48 GB of VRAM:&lt;/strong&gt; 27B dense models like Qwen3.8-27B, at 4-bit (about 13.5 GB) or 8-bit (about 27 GB). BF16 needs about 54 GB, more than a single 48 GB card holds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Server or multi-machine setup:&lt;/strong&gt; Muse Spark Open and DeepSeek V4.1 Flash 1017. Even DeepSeek's 552B backbone alone is about 276 GB at 4-bit.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For models bigger than your VRAM, llama.cpp offers CPU+GPU hybrid inference to partially accelerate them.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Do MoE models need less memory because only some parameters are active?
&lt;/h3&gt;

&lt;p&gt;No. Active parameters cut compute per token, but every expert must still be loaded. Muse Spark Open activates 42B of its 816B parameters, yet Meta pairs two Mac Studios so they "can hold an entire model." The MoE win is speed: Meta notes each token reads only a small part of the model from memory.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is an open-weight model free to use commercially?
&lt;/h3&gt;

&lt;p&gt;It depends on the license, not the word "open." MIT and Apache 2.0 models like DeepSeek V4.1 Flash 1017 and Qwen3.8-27B grant broad rights, including distributing copies, under their notice terms. Muse Spark Open allows a lot too, with attribution, a usage policy and the revenue clause above.&lt;/p&gt;

&lt;h2&gt;
  
  
  The takeaway
&lt;/h2&gt;

&lt;p&gt;Before you download anything, multiply total parameters by bits per weight, divide by 8, and add headroom for the context length you actually need. If that fits your RAM or VRAM, read the license next, before the model ends up in a product.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/" rel="noopener noreferrer"&gt;Meta launches an open-weight AI model that runs on Macs (TechCrunch)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.engadget.com/2287833/meta-built-a-huge-open-weight-ai-model-for-macs-but-youll-need-1tb-of-ram-to-run-it/" rel="noopener noreferrer"&gt;Meta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run it (Engadget)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://research.meta.ai/blog/muse-spark-open" rel="noopener noreferrer"&gt;Introducing Muse Spark Open (Meta AI Research)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B" rel="noopener noreferrer"&gt;Muse-Spark-Open-816B-A42B model card and Muse Model License (Hugging Face)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://api-docs.deepseek.com/news/news261017" rel="noopener noreferrer"&gt;DeepSeek-V4.1-Flash-1017 release note (DeepSeek API Docs)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017" rel="noopener noreferrer"&gt;DeepSeek-V4.1-Flash-1017 model card (Hugging Face)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash" rel="noopener noreferrer"&gt;DeepSeek-V4.1-Flash model card (Hugging Face)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/ibm-granite/granite-switch-4.1-8b" rel="noopener noreferrer"&gt;Granite Switch 4.1 8B model card (Hugging Face)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/ibm-granite/granite-switch-4.1-8b-GGUF" rel="noopener noreferrer"&gt;granite-switch-4.1-8b GGUF (Hugging Face)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/Qwen/Qwen3.8-27B" rel="noopener noreferrer"&gt;Qwen3.8-27B model card (Hugging Face)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://opensource.org/ai/open-source-ai-definition" rel="noopener noreferrer"&gt;The Open Source AI Definition 1.0 (Open Source Initiative)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://opensource.org/ai/open-weights" rel="noopener noreferrer"&gt;Open Weights (Open Source Initiative)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/docs/transformers/llm_tutorial_optimization" rel="noopener noreferrer"&gt;Optimizing LLMs for Speed and Memory (Hugging Face Transformers docs)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/" rel="noopener noreferrer"&gt;Introducing NVFP4 for Efficient and Accurate Low-Precision Inference (NVIDIA Developer Blog)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/ggml-org/llama.cpp" rel="noopener noreferrer"&gt;llama.cpp README (GitHub)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/LICENSE" rel="noopener noreferrer"&gt;DeepSeek-V4.1-Flash-1017 LICENSE (MIT)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.apache.org/licenses/LICENSE-2.0" rel="noopener noreferrer"&gt;Apache License, Version 2.0 (Apache Software Foundation)&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;This article was written with AI assistance and reviewed by the author.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>opensource</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>How to write your first Agent Skill: SKILL.md explained with a working example</title>
      <dc:creator>Nirali</dc:creator>
      <pubDate>Sat, 24 Oct 2026 07:02:14 +0000</pubDate>
      <link>https://dev.to/nirali_toolbit/how-to-write-your-first-agent-skill-skillmd-explained-with-a-working-example-afp</link>
      <guid>https://dev.to/nirali_toolbit/how-to-write-your-first-agent-skill-skillmd-explained-with-a-working-example-afp</guid>
      <description>&lt;p&gt;One folder and one Markdown file can teach an AI coding agent a repeatable task. Here's the format, the rules that trip people up, and a skill you can copy today.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;In short:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A skill is a folder with one required file, &lt;code&gt;SKILL.md&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;The &lt;code&gt;description&lt;/code&gt; field decides when an agent uses the skill, so write it first.&lt;/li&gt;
&lt;li&gt;Validate the name before you share it, because some tools won't load a skill with an invalid name.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What an Agent Skill is
&lt;/h2&gt;

&lt;p&gt;An Agent Skill is a folder that teaches an AI agent how to do one specific job. At minimum it holds a single &lt;code&gt;SKILL.md&lt;/code&gt; file: YAML frontmatter on top, plain Markdown instructions below (&lt;a href="https://agentskills.io/specification" rel="noopener noreferrer"&gt;Agent Skills specification&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;The format started at Anthropic and was published as an open standard in December 2025, with tools such as Cursor, GitHub, VS Code, goose and OpenCode listed as early adopters (&lt;a href="https://simonwillison.net/2025/Dec/19/agent-skills" rel="noopener noreferrer"&gt;Simon Willison&lt;/a&gt;). Editors like Zed now load skills from an &lt;code&gt;.agents/skills/&lt;/code&gt; folder (&lt;a href="https://zed.dev/docs/ai/skills.md" rel="noopener noreferrer"&gt;Zed docs&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;The appeal is simple. You write instructions once, and any compatible agent can pick them up when a task matches.&lt;/p&gt;

&lt;h2&gt;
  
  
  How agents load a skill
&lt;/h2&gt;

&lt;p&gt;Skills are designed to stay out of the way until they're needed. The spec calls this progressive disclosure, and it works in three steps (&lt;a href="https://agentskills.io/specification" rel="noopener noreferrer"&gt;spec&lt;/a&gt;):&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;At startup the agent reads only each skill's &lt;code&gt;name&lt;/code&gt; and &lt;code&gt;description&lt;/code&gt;, about 100 tokens per skill.&lt;/li&gt;
&lt;li&gt;When a task matches the description, it loads the full &lt;code&gt;SKILL.md&lt;/code&gt; body. The spec recommends keeping that under 5,000 tokens.&lt;/li&gt;
&lt;li&gt;Files in &lt;code&gt;scripts/&lt;/code&gt;, &lt;code&gt;references/&lt;/code&gt; or &lt;code&gt;assets/&lt;/code&gt; load only when the instructions point to them.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;That's why the &lt;code&gt;description&lt;/code&gt; field matters more than anything else in the file. It is the only thing the agent sees when deciding whether to use your skill.&lt;/p&gt;

&lt;h2&gt;
  
  
  The folder layout
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;changelog-writer/
├── SKILL.md
└── references/
    └── STYLE.md
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Only &lt;code&gt;SKILL.md&lt;/code&gt; is required. The &lt;code&gt;scripts/&lt;/code&gt;, &lt;code&gt;references/&lt;/code&gt; and &lt;code&gt;assets/&lt;/code&gt; folders are optional conventions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the folder goes
&lt;/h2&gt;

&lt;p&gt;Each tool reads skills from its own location, so check your editor's docs. In Zed, for example, skills live in &lt;code&gt;~/.agents/skills/&lt;/code&gt; for every project or in &lt;code&gt;.agents/skills/&lt;/code&gt; inside one project. Each skill must sit directly in that folder, with no nested groups, and project skills only load once you've marked the project as trusted (&lt;a href="https://zed.dev/docs/ai/skills.md" rel="noopener noreferrer"&gt;Zed docs&lt;/a&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  The frontmatter rules that trip people up
&lt;/h2&gt;

&lt;p&gt;Two fields are required (&lt;a href="https://agentskills.io/specification" rel="noopener noreferrer"&gt;spec&lt;/a&gt;):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;name&lt;/code&gt;&lt;/strong&gt;: 1 to 64 characters, lowercase letters, numbers and hyphens only. It can't start or end with a hyphen, can't contain &lt;code&gt;--&lt;/code&gt;, and must match the folder name.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;description&lt;/code&gt;&lt;/strong&gt;: 1 to 1,024 characters. It should say what the skill does and when to use it, with keywords an agent can match.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Optional fields include &lt;code&gt;license&lt;/code&gt;, &lt;code&gt;compatibility&lt;/code&gt; (up to 500 characters, only if the skill needs a specific environment), &lt;code&gt;metadata&lt;/code&gt; and an experimental &lt;code&gt;allowed-tools&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The most common mistakes are an uppercase name, a folder name that doesn't match, and a vague description like "Helps with changelogs". In Zed, for example, a skill with an invalid name fails to load and shows an error (&lt;a href="https://zed.dev/docs/ai/skills.md" rel="noopener noreferrer"&gt;Zed docs&lt;/a&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  A complete example: changelog-writer
&lt;/h2&gt;

&lt;p&gt;This skill turns recent commits into release notes in a consistent style.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="nn"&gt;---&lt;/span&gt;
&lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;changelog-writer&lt;/span&gt;
&lt;span class="na"&gt;description&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Writes&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;release&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;notes&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;from&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;recent&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;git&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;commits&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;in&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;Keep&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;a&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;Changelog&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;style.&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;Use&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;when&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;the&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;asks&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;for&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;a&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;changelog,&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;release&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;notes,&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;or&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;a&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;summary&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;of&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;what&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;changed&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;since&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;the&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;last&lt;/span&gt;&lt;span class="nv"&gt; &lt;/span&gt;&lt;span class="s"&gt;tag."&lt;/span&gt;
&lt;span class="nn"&gt;---&lt;/span&gt;

&lt;span class="gh"&gt;# Changelog writer&lt;/span&gt;
&lt;span class="p"&gt;
1.&lt;/span&gt; Find the latest tag with &lt;span class="sb"&gt;`git describe --tags --abbrev=0`&lt;/span&gt;.
&lt;span class="p"&gt;2.&lt;/span&gt; List commits since that tag with &lt;span class="sb"&gt;`git log &amp;lt;tag&amp;gt;..HEAD --oneline`&lt;/span&gt;.
&lt;span class="p"&gt;3.&lt;/span&gt; Group each commit under Added, Changed, Fixed or Removed.
&lt;span class="p"&gt;4.&lt;/span&gt; Rewrite each line for users, not developers: say what changed for them.
&lt;span class="p"&gt;5.&lt;/span&gt; Skip merge commits and dependency bumps unless they fix a security issue.

Follow the tone and formatting rules in &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nv"&gt;the style guide&lt;/span&gt;&lt;span class="p"&gt;](&lt;/span&gt;&lt;span class="sx"&gt;references/STYLE.md&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;.

&lt;span class="gu"&gt;## Example&lt;/span&gt;

Input commit: &lt;span class="sb"&gt;`fix: handle empty CSV upload without crashing`&lt;/span&gt;
Output line: &lt;span class="sb"&gt;`Fixed: uploading an empty CSV file no longer causes an error.`&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The style guide sits in &lt;code&gt;references/STYLE.md&lt;/code&gt;, so it loads only when the agent actually writes a changelog.&lt;/p&gt;

&lt;h2&gt;
  
  
  Keep it small
&lt;/h2&gt;

&lt;p&gt;The spec gives three practical limits worth following (&lt;a href="https://agentskills.io/specification" rel="noopener noreferrer"&gt;spec&lt;/a&gt;):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Keep &lt;code&gt;SKILL.md&lt;/code&gt; under 500 lines and move long material into reference files.&lt;/li&gt;
&lt;li&gt;Keep file references one level deep from &lt;code&gt;SKILL.md&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Reference files by relative path from the skill folder.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Validate before you share
&lt;/h2&gt;

&lt;p&gt;The standard ships a reference library, &lt;code&gt;skills-ref&lt;/code&gt;, that checks your frontmatter and naming:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;skills-ref validate ./changelog-writer
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Run it before committing a skill to a shared repo. Some agents simply skip a skill whose name breaks the rules, so it's easier to catch problems here.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check that your agent actually uses it
&lt;/h2&gt;

&lt;p&gt;A valid skill can still sit unused if its description doesn't match how people ask for the task. In Zed you can test both paths: type &lt;code&gt;/&lt;/code&gt; and pick the skill to load it by hand, or describe the task in plain words and see whether the agent reaches for it on its own (&lt;a href="https://zed.dev/docs/ai/skills.md" rel="noopener noreferrer"&gt;Zed docs&lt;/a&gt;). If it never triggers by itself, rewrite the description before you touch the steps.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;Start with one task your team repeats every week, such as release notes, PR descriptions or test scaffolding. Write the description first, then the steps.&lt;/p&gt;

&lt;p&gt;Still choosing an editor to run your skills in? This &lt;a href="https://www.toolbit.ai/blog/cursor-vs-windsurf-which-agentic-ide-solves-m" rel="noopener noreferrer"&gt;Cursor vs Windsurf comparison&lt;/a&gt; puts both through a repository-wide refactor.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Disclosure: I work on Toolbit, which published the comparison linked above. This article was created with the help of AI and reviewed by the author for accuracy.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>tutorial</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
