Archivist

The canonical Claude ecosystem reference

Auto-synced official docs and curated Anthropic team insights, built for deterministic local lookup via the Claude Code CLI — no RAG, no retrieval blind spots.

• Documentation
• Using the Docs
  • Recommended — Claude Code CLI (deterministic)
    • Companion skill — ask-archivist
  • Alternative — Claude Project (probabilistic, with limits)
• Automated Updates
• License

Auto-updated mirror of the official Claude documentation, refreshed every 12h via GitHub Actions. Structure follows Claude Project Knowledge Base best practices.

Folders mirror the official upstream sitemap sections 1:1, so navigation matches platform.claude.com and code.claude.com.

There are two ways to consume this knowledge base. They are not equivalent — pick based on the precision you need.

git clone https://github.com/coroboros/archivist.git
cd archivist
claude

Claude Code reads files directly with Grep (full-text, deterministic) and Read (full file content). Every fact in the docs is reachable; no fact is silently dropped. Cross-references resolve dynamically because Claude can follow links across the entire mirror in one session.

This is the right path when you need surgical accuracy — writing technical docs, reviewing implementation choices against current Anthropic guidance, refining skills, cross-referencing API ↔ Code ↔ Platform guides.

Use docs/INDEX.md for cross-section topic navigation, then Grep/Read from there.

This repo ships claude/skills/ask-archivist/SKILL.md, a companion skill that wraps the deterministic lookup workflow (INDEX Grep → section Grep → full-file Read → cite). Install it once, then ask questions in natural language — the skill auto-detects the mirror and cites the files it reads.

# From an archivist clone
mkdir -p ~/.claude/skills/ask-archivist
cp claude/skills/ask-archivist/SKILL.md ~/.claude/skills/ask-archivist/

# Point the skill at any mirror location (optional — defaults to $HOME/archivist or the current repo)
export ARCHIVIST_PATH="$PWD"

Why this over the built-in claude-code-guide subagent? Both answer the same kinds of questions, but on opposite ends of the precision/freshness tradeoff:

Use ask-archivist by default for precision work. Fall back to claude-code-guide when the change is < 12h old or the question is about runtime behavior / community patterns the docs don't cover — the skill explicitly names this escape-hatch so you keep the user in control.

1. Create a Claude Project
2. Connect this repo as a GitHub data source (or upload docs/ folders)
3. Use the system prompt at claude/archivist-project/system-prompt.xml

Convenient for mobile, non-technical users, or quick queries away from a terminal — but understand the tradeoff before relying on it for precise work:

• Claude Projects use RAG (retrieval-augmented generation) once content exceeds the context window: docs are chunked, embedded, and retrieved by semantic similarity to your query.
• RAG is probabilistic: the relevant chunk may not surface in the top-k results, especially when your query phrasing differs from the doc's wording. Misses are silent — the model fills gaps with training-data approximations.
• Concrete example: asked about Claude Code worktrees, Claude Desktop has been observed to suggest manual git worktree setup, missing the native claude --worktree flag — even though it appears in 27 files of this mirror.

Full pedagogical write-up: claude/archivist-project/README.md.

For high-precision work (regulated content, technical writing, audit-grade research), use the CLI path above.

The GitHub Actions workflow runs every 12h:

1. Fetches latest Claude docs from Anthropic sitemaps
2. Generates markdown files with YAML frontmatter
3. Rebuilds README indexes and the cross-section docs/INDEX.md
4. Auto-commits, tags, creates GitHub releases

Scripts run on Bun. To update docs locally: pnpm update (requires bun and pnpm).

Safety guard: Aborts if > 25% of files in any docs folder are deleted.

MIT