docs

name	docs
description	Search and access Anthropic documentation covering Claude Code CLI, API, Agent SDK, and more. Supports natural language queries, content search, and direct topic lookup. Trigger on: /docs, documentation questions about Claude Code, Claude API, Agent SDK, prompt engineering, MCP, hooks, skills, tool use, streaming, batch processing, extended thinking, or any Anthropic platform feature.
allowed-tools	Bash Read Glob Grep AskUserQuestion

Overview

This skill provides AI-powered search and access to a locally mirrored copy of Anthropic's official documentation. The documentation lives at ~/.claude-code-docs and covers ~1,700 paths across four categories.

When a user asks about Anthropic documentation, use the search tools and reference files described below to find the relevant content, read it, and synthesize an answer. Do not guess — always read the actual documentation before answering.

Domain Concept Map

Documentation sources: Two Anthropic domains are mirrored:

• code.claude.com — Claude Code CLI documentation
• platform.claude.com — Everything else: API, Agent SDK, guides

Four categories organize the paths (approximate counts as of 2026-06):

Category	User Label	Paths	Covers
claude_code	Claude Code CLI	~43	CLI setup, hooks, skills, MCP, memory, plugins, settings, sub-agents
api_reference	Claude API	~1,460	Messages API, models, batches, files, admin, multi-language SDKs (Python/TS/Go/Java/Kotlin/Ruby), Agent SDK
core_documentation	Claude Documentation	~200	Prompt engineering, tool use, vision, streaming, extended thinking, evaluation, release notes, resources
uncategorized	Other	~1	Paths not yet assigned to a category

Agent SDK paths live within api_reference but are labeled "Claude Agent SDK" for users. They cover: overview, Python/TypeScript SDKs, sessions, skills, subagents, MCP, plugins, structured outputs, and more.

File naming convention: Documentation files use double underscores for path separators:

• docs__en__hooks.md — Claude Code CLI page /docs/en/hooks
• en__docs__claude-code__hooks.md — Alternate format for the same page
• en__api__messages__create.md — API reference page /en/api/messages/create

How to Search

Follow this workflow when handling documentation queries:

Step 1: Sync the latest docs (MANDATORY — always do this first)

Before searching or reading anything, pull the latest documentation. This is the first action for every docs request — it guarantees fresh content and clones the mirror on first use. There is no search index to build (search runs over the live files). Run exactly this:

DOCS="$HOME/.claude-code-docs"
if [ -d "$DOCS/.git" ]; then
  git -C "$DOCS" pull --ff-only --quiet 2>/dev/null || true
else
  git clone --quiet https://github.com/seanGSISG/claude-code-docs.git "$DOCS"
fi
# Refresh the runtime helper from tracked source (used for direct topic reads)
[ -f "$DOCS/scripts/claude-docs-helper.sh" ] && { cp "$DOCS/scripts/claude-docs-helper.sh" "$DOCS/claude-docs-helper.sh" 2>/dev/null; chmod +x "$DOCS/claude-docs-helper.sh" 2>/dev/null; }
# Best-effort: ensure ripgrep for the standalone helper (the Grep tool already
# bundles ripgrep, so this is optional; grep is the universal fallback).
if ! command -v rg >/dev/null 2>&1; then
  if command -v winget >/dev/null 2>&1; then winget install --silent --accept-source-agreements --accept-package-agreements BurntSushi.ripgrep.MSVC >/dev/null 2>&1 || true
  elif command -v brew >/dev/null 2>&1; then brew install ripgrep >/dev/null 2>&1 || true
  elif command -v apt-get >/dev/null 2>&1; then sudo -n apt-get install -y ripgrep >/dev/null 2>&1 || true
  fi
fi
 

Proceed to the next step regardless of the sync result — if the network is down, the existing local mirror is still usable.

Step 2: Analyze Intent

Extract from the user's query:

• Keywords — the specific concepts they want (e.g., "hooks", "extended thinking", "batch API")
• Product context — if they specify one (e.g., "in the agent sdk", "cli hooks", "api rate limits")
• Query type — how-to, reference lookup, comparison, discovery

Step 3: Locate the doc(s)

Search the live files with Claude Code's built-in tools — they use bundled ripgrep, so they are fast and return only compact results (paths/snippets), keeping token use low. Search the docs directory ~/.claude-code-docs/docs (use the absolute $HOME path when calling the tools).

• Content search — the Grep tool over the docs directory:
  • Find candidate files: output_mode: "files_with_matches".
  • Rank by relevance: output_mode: "count" and prefer files with more matches.
  • Disambiguate without opening files: output_mode: "content", -C: 2, small head_limit.
• Topic / path lookup — the Glob tool (filenames encode the path):
  • **/*<topic>*.md over the docs directory — e.g. hooks → claude-code__hooks.md, docs__en__hooks.md.

Filenames map directly to doc paths, so a filename match is usually enough to pick the right doc. There is no index — results are always current.

Shell fallback (e.g., when driving the helper standalone, outside the agent):

~/.claude-code-docs/claude-docs-helper.sh --search-content "<keywords>"   # ripgrep, grep fallback
~/.claude-code-docs/claude-docs-helper.sh --search "<topic>"              # path match
~/.claude-code-docs/claude-docs-helper.sh <topic>                         # direct topic read

See references/search-guide.md for detailed search tool usage.

Step 4: Decide — Synthesize or Ask

Check which product categories the results span:

• Same category (e.g., all Claude Code CLI) → Read all matching docs silently, synthesize a unified answer. Never ask "which doc do you want?" when results are in the same product context.
• Multiple categories (e.g., CLI + API + Agent SDK) → Ask the user which product context using AskUserQuestion with user-friendly product labels.

See references/category-map.md for the full category-to-label mapping and disambiguation rules.

Step 5: Read and Present

1. Read the matching documentation files using their file paths from the search results
2. Extract sections relevant to the user's question
3. Synthesize a unified answer combining insights from all sources
4. Cite all sources at the end with official documentation URLs

Reference Navigation

Load the reference file matching the topic before answering detailed questions about search mechanics or category routing.

Topic	Reference File	Key Contents
How to use search tools, filename conventions, Python fallback, direct doc reading	references/search-guide.md	Search commands, output formats, file naming patterns, graceful degradation
Product categories, user-facing labels, disambiguation rules, when to ask vs synthesize	references/category-map.md	Category-to-label map, path patterns, cross-context resolution strategy

Cross-Reference Guide

Some queries span multiple reference files or require special handling:

Question Pattern	Action
"How do I use X in agent sdk?"	Filter search to agent-sdk paths, read all matches, synthesize
"What's the difference between X and Y?"	Search for both terms, read docs for each, present comparison
"Show me all docs about X"	Run path search, present grouped by product category
"hooks" (ambiguous — CLI hooks vs Agent SDK hooks)	Search content, check categories — if split across products, ask user
Direct topic name (e.g., "mcp", "memory")	Try direct file read first: ~/.claude-code-docs/docs/docs__en__<topic>.md
Freshness check (-t)	Run: ~/.claude-code-docs/claude-docs-helper.sh -t
"what's new"	Run: ~/.claude-code-docs/claude-docs-helper.sh "what's new"
"update the docs" / "pull latest docs"	Run the Step 1 sync block above (or tell the user to run the /docs-update command, which does the same thing).

Keeping Docs Current

There is no background hook and no search index — syncing is explicit and always visible. The mirror is refreshed by the mandatory Step 1 sync block at the start of this workflow, so every docs request pulls the latest content; search then runs over the live files, so results are always current. The /docs-update command runs the same sync on demand without a search.

Key Commands Quick Reference

# Content search (returns JSON with product context)
~/.claude-code-docs/claude-docs-helper.sh --search-content "extended thinking"

# Path search (returns ranked path matches)
~/.claude-code-docs/claude-docs-helper.sh --search "hooks"

# Direct topic lookup
~/.claude-code-docs/claude-docs-helper.sh hooks

# List all available docs
ls ~/.claude-code-docs/docs/*.md | sed 's/.*\///' | sed 's/\.md$//'

# Check freshness and sync
~/.claude-code-docs/claude-docs-helper.sh -t

# What's new
~/.claude-code-docs/claude-docs-helper.sh "what's new"

# Installation status
~/.claude-code-docs/claude-docs-helper.sh --status

# Fallback search (no helper script)
grep -ril "keyword" ~/.claude-code-docs/docs/ | head -20

Important Caveats

• Documentation is a mirror, not the source. Always note that content comes from Anthropic's official documentation. Include official URLs when citing.
• Two base URLs: Claude Code CLI pages are at code.claude.com/docs/en/<page>. Everything else is at platform.claude.com/<path>.
• No search index — content search runs over the live files via the Grep tool (bundled ripgrep) or the helper's --search-content (ripgrep, with grep as fallback). Always current; nothing to rebuild.
• Helper script requires Python 3.9+ for enhanced features. Basic topic lookup and freshness checks work without Python. Content and path search need Python.
• ~1,530 files, ~1,702 manifest paths. The manifest tracks more paths than there are files; some are tracked but not downloadable (expected). Exact counts grow as Anthropic publishes docs — run ~/.claude-code-docs/claude-docs-helper.sh --status for live numbers.