claude-dev-skill

Unofficial, auto-updating agent skill built from the claude.dev engineering blog

imageLicense: MIT

Anthropic's developers publish playbooks and engineering write-ups on claude.dev: how to pick an effort level, what a task costs, how to keep the prompt cache warm, how to write skills and mods, how to design evals. This skill turns them into short working notes an agent can load in the middle of a task, links every note to its posts, and reads the original post from claude.dev when the detail matters.

A scheduled job checks the site every hour. When a post or mod is added or changed, it refreshes the index, rewrites the affected notes and publishes a new release once an independent check passes.

What's inside

skills/claude-dev-skill/

Ask your agent, for example:

• "Pick an effort level for this task using the claude.dev guidance."
• "What does claude.dev say about structuring this prompt for caching?"
• "Are there claude.dev posts newer than this skill's notes?"

Install

npx · skills

npx skills add buidangminh23/claude-dev-skill

Claude Code · plugin (run inside Claude Code)

/plugin marketplace add buidangminh23/claude-dev-skill
/plugin install claude-dev-skill@claude-dev-skill

Codex · marketplace

codex plugin marketplace add buidangminh23/claude-dev-skill

Gemini CLI · extension

gemini extensions install https://github.com/buidangminh23/claude-dev-skill

Manual

Download claude-dev-skill-vX.Y.Z.zip and SHA256SUMS.txt from Releases, check the hash, and copy skills/claude-dev-skill into ~/.claude/skills/ or ~/.codex/skills/. The claude-dev-skill-plugin-vX.Y.Z.zip asset has plugin.json at its root for clients that install plugin archives.

The whatsnew command needs Node.js 22 or newer and network access. Without them, the skill tells the agent to compare https://claude.dev/llms.txt with references/index.md instead.

How it stays current

every hour   detect      conditional GETs to posts.json, llms.txt, /mods/, /terms/ and each post's .md
             index       regenerate index.md and mods-catalog.md, record content hashes
on change    distill     read-only job: the Claude Code GitHub Action edits notes for up to 3 new or
                         edited posts and hands them over as an artifact
             guard       publishing job, on its own checkout: own words (no 12-word run shared with
                         any post, at most 6% shared 6-word sequences), allowed links only, no risky
                         commands, no deleted notes, a line budget, frozen parts of SKILL.md, size
                         limits, no secrets or personal data
             release     bump the version, tag, upload to a draft, re-download and verify, mark latest
anytime      attention   one issue tracks anything that needs a person, and closes itself when clear

Posts that cannot be distilled yet stay in a pending list and are retried every six hours. Agents still find them through the generated index and whatsnew. Distillation pauses when the claude.dev Terms of Use change, until someone reviews them and updates data/policy.json.

Security model

The model reads text from the web, so it never holds a credential that can change this repository. The distill job runs with a read-only token, may edit only the notes and the description, note table and principles of SKILL.md, cannot use web tools, and can run only the guard; its shell commands run inside bubblewrap with the job's credentials scrubbed from their environment. The publishing job never executes anything from the artifact: it copies notes with expected names, fetches the posts itself and runs the guard from main before anything is committed. The rest of SKILL.md, the scripts and the workflows change only through commits by people. Actions are pinned to commit SHAs and updated by Dependabot.

One-time setup for maintainers

1. Create a token with claude setup-token on your machine, then store it as a repository secret named CLAUDE_CODE_OAUTH_TOKEN (gh secret set CLAUDE_CODE_OAUTH_TOKEN -R buidangminh23/claude-dev-skill). Only the read-only distill job uses it. Without it, the index and releases still update and the attention issue lists the posts waiting for notes.
2. Optional: PERSONAL_WEB_DISPATCH_TOKEN, a fine-grained token with Contents read and write on the Personal-Web repository, so the skill card on the site refreshes right after a release.

Development

npm test                                  # unit tests
node scripts/detect.mjs                   # dry run against claude.dev
node scripts/fetch-sources.mjs .cache/sources
node scripts/guard.mjs --sources .cache/sources
node scripts/release.mjs check            # versions, manifests, changelog, web card numbers

Not affiliated with Anthropic

This project is independent. It is not affiliated with, endorsed by, or sponsored by Anthropic. "Claude" and "Anthropic" are trademarks of Anthropic, PBC. Articles on claude.dev belong to Anthropic; this repository does not copy or republish them. It links to them, quotes at most brief excerpts with attribution, and summarizes in its own words. Code samples on claude.dev are MIT-licensed by Anthropic unless a sample says otherwise.

If you hold rights in anything here and want it changed or removed, open an issue in this repository. Distillation can be paused at once, and affected releases will be withdrawn.

License

MIT for the contents of this repository. See LICENSE. The Claude Code mark (the project icon) and the Codex mark in assets/ are trademarks of Anthropic, PBC and OpenAI; they are reproduced unchanged and do not mean either company endorses this project. See assets/SOURCES.md and THIRD_PARTY_NOTICES.md.