claude-docs

name	claude-docs
description	Ground Claude Code configuration and feature work — and any work on Anthropic's products — in the latest OFFICIAL docs instead of training knowledge, which drifts because these ship constantly. Three doc hosts. Claude Code (code.claude.com) — skills, hooks, subagents, slash/custom commands, MCP servers, settings.json, permissions, plugins, statusline, output styles, memory/CLAUDE.md, the CLI, the Agent SDK. The Anthropic API / developer platform (platform.claude.com) — Messages API, models, prompt caching, tool use, batch, files, citations, prompt engineering, Managed Agents, Agent Skills. The Help Center (support.claude.com) — account, billing, SSO/JIT/SCIM, Connectors, Claude in Chrome, how-to guides, release notes. Use when building, editing, auditing, reviewing, or debugging any of the above — including authoring or scaffolding a skill, subagent, hook, slash command, or plugin — or answering "how does Claude Code / the Claude API do X" or "what's the latest way to configure X". Always prefer this over guessing.
when_to_use	Use to GROUND work in current official docs — look up the authoritative page, verify a field or behavior, cite it. For actually building, debugging, or migrating a Claude API / Anthropic SDK app (prompt caching, model version bumps, tool-use wiring) prefer the claude-api skill; this skill supplies the docs, not the implementation.

Claude docs (live)

Fetch current official Anthropic documentation on demand. Never answer a Claude Code, Claude API, or Anthropic product question from memory when the live docs can confirm it — these products ship often and training data drifts.

The three hosts (route by domain of the question)

Pick the authoritative host first; don't blend them.

Host	Authoritative for	Index (llms.txt)
code.claude.com/docs	Claude Code — CLI, skills, hooks, subagents, slash/custom commands, MCP servers, settings.json, permissions, plugins, statusline, output styles, memory/CLAUDE.md, Agent SDK	https://code.claude.com/docs/llms.txt (~30 KB)
platform.claude.com	Anthropic API / dev platform — Messages API, models, prompt caching, tool use, batch, files, citations, prompt engineering, Managed Agents, Agent Skills, build-with-claude	https://platform.claude.com/llms.txt (~167 KB)
support.claude.com	Help Center — account, billing, SSO/JIT/SCIM, Connectors, Claude in Chrome, how-to guides, release notes (multilingual)	https://support.claude.com/llms.txt (~172 KB)

Routing rules:

• A Claude Code config/feature question → code.claude.com is authoritative. If a support.claude.com how-to article and a code.claude.com reference page disagree, the code.claude.com page wins — support is end-user help, not the spec. ("Skills" appears on all three hosts; for Claude Code skill-building, use code.claude.com.)
• Building or debugging an API/SDK app (calling Claude, caching, tool use, models) → platform.claude.com.
• Account / billing / SSO / "how do I use X in the app" / release notes → support.claude.com. Its llms.txt is multilingual — stay in the ## English section.

The docs contract (stable; everything else is fetched live)

All three are Mintlify-hosted and share these conventions. Do not memorize page slugs — they change.

1. Raw markdown for ANY page — append .md. Append .md to any bare docs/article URL on these hosts to get clean markdown with no nav chrome: …/en/hooks → …/en/hooks.md, …/articles/123-foo → …/articles/123-foo.md. The bare HTML URL is a multi-MB SPA shell; the .md is small, pure content. Never double-append (.md.md falls back to HTML). Verified on code, platform, and support alike.

2. Index per host: the llms.txt files in the table above — every page with a one-line description and its .md URL. The source of truth for "which page covers what." Start here only when you don't already have a URL.

3. Aliases: docs.claude.com and docs.anthropic.com both 301-redirect to platform.claude.com. Use platform.claude.com canonically; don't hardcode the aliases.

4. Everything (last resort): each host has an llms-full.txt (entire docs concatenated). Sizes differ wildly — code ≈ 4 MB, platform ≈ 71 MB. Use only for full-text search when the index can't locate a term. Do not WebFetch it (it summarizes MBs through a small model every call) and never casually curl platform's whole 71 MB — curl + grep it as a search index, caching once to /tmp if grepping several terms this session:

   # locate which page mentions a term (-i case-insensitive, -n line numbers)
   curl -s https://code.claude.com/docs/llms-full.txt | grep -ni 'disable-model-invocation'
   # several terms this session? fetch once (CDN caches ~daily), then grep /tmp:
   curl -s https://code.claude.com/docs/llms-full.txt -o /tmp/cc-llms-full.txt
   grep -ni -B2 -A8 'your-term' /tmp/cc-llms-full.txt
   

   Only grep output enters context, never the raw file. No manual cleanup — the OS reaps /tmp, and the cached copy warms the next lookup.

Workflow

Delegate the lookup to a subagent so the index and page bodies stay out of the main context — doing it inline floods the transcript and burns the main context window. Spawn a subagent (the Agent tool, subagent_type: "Explore" or general-purpose) that fetches in its own isolated context and returns only the distilled answer plus the page name to cite.

Always curl — not WebFetch — because the pages are static markdown on a CDN: instant, free, not rate-limited. WebFetch throttles (the attempt N/10 retries) and returns a lossy small-model summary. If curl is ever blocked (unexpected for these static files), fall back to firecrawl scrape "<url>" (costs credits). Do not use context7 — it indexes libraries, not these live docs.

The subagent's steps:

1. Already have a doc URL? Fast-path it. curl -s <url>.md (append .md to the bare URL, or use it as-is if it already ends in .md) and read it. No index round-trip — the index is only for discovery.
2. Otherwise route, then discover. Pick the host from the table above. curl -s <host-llms.txt> and find the entry whose title/description matches. Take its .md URL.
3. Read the page. curl -s <that-.md-url> and read the relevant section.
4. Fallbacks, in order: index has no clear match → curl + grep that host's llms-full.txt (see contract), then resolve its .md; a .md 404s → curl the bare path (HTML still works).
5. Return shape. Reply with just (a) the answer to the specific question and (b) the page name(s) for citation — not the index, not the full page body.

Then cite it in your own work — name the page the subagent reported, e.g. "per code.claude.com … hooks.md → PreToolUse" — so the user can verify.

Examples

Build a new Claude Code skill, want official guidance. Spawn a subagent: "Route via https://code.claude.com/docs/llms.txt, find the best-practices / skills page, curl its .md, and summarize what the Create skills section recommends. Return only that plus the page name." It resolves code.claude.com/docs/en/best-practices.md, reads #create-skills, returns the summary. Apply and cite.

Someone pasted https://support.claude.com/en/articles/12512198-how-to-create-custom-skills. Fast-path it — curl -s …/12512198-how-to-create-custom-skills.md returns clean markdown, no index needed. (That's the end-user how-to; for the authoritative Claude Code skill spec, prefer code.claude.com.)

Tuning prompt caching in an API app. Route to platform.claude.com: curl its llms.txt, find the prompt-caching page, curl that .md.

Notes

• Prefer the index→page path (or the URL fast-path) over llms-full.txt; it's far cheaper in tokens.
• This skill complements the built-in claude-code-guide agent and the vendored claude-api skill — use this to ground your own build/config work in current docs.