Keep CLAUDE.md Accurate as Code Evolves with Caliber

Hand-written CLAUDE.md files go stale the moment you refactor — agents hallucinate paths that no longer exist and advise from yesterday's architecture. Caliber generates and maintains your AI context files, then keeps them in sync as the code moves, for Claude Code, Cursor, Codex, OpenCode, and GitHub Copilot.

Why This Skill Matters

Caliber treats agent config as code. A deterministic score — no LLM, no API calls — cross-references your config files against the actual project filesystem across six categories: Files & Setup (25 points), Quality (25), Grounding (20), Accuracy (15), Freshness & Safety (10), and Bonus (5). The workflow mirrors code review: score first, propose changes as a diff, review each one, back up originals to .caliber/backups/, and undo with caliber undo. If your existing config already scores 95+, Caliber skips full regeneration and applies targeted fixes to the failing checks. The repo (caliber-ai-org/ai-setup) has 1,267 stars on the SkillMap leaderboard.

Installation

Requires Node.js 20 or later.

npx @rely-ai/caliber bootstrap

Bootstrap takes about two seconds and installs the /setup-caliber skill — the README describes it as 100% local, with no LLM calls and no code sent anywhere. Then start a Claude Code or Cursor CLI session in your terminal and type:

/setup-caliber

Your agent detects the stack, generates tailored configs for every platform your team uses, sets up pre-commit hooks, and enables continuous sync. Not on Claude Code or Cursor? caliber init runs the same setup as a CLI wizard.

Real Workflow: Bootstrap, Score, and Watch the Loop

1. Run npx @rely-ai/caliber bootstrap, then /setup-caliber in your session. The agent generates per-platform configs — CLAUDE.md plus CALIBER_LEARNINGS.md and .claude/skills/*/SKILL.md for Claude Code, .cursor/rules/*.mdc for Cursor, AGENTS.md for Codex and OpenCode, .github/copilot-instructions.md for Copilot.
2. Score your setup deterministically, and compare across branches:

caliber score --compare main

3. Let the loop run. Pre-commit hooks trigger caliber refresh, which analyzes committed, staged, and unstaged changes and updates the config files to match. New teammates get nudged to bootstrap on their first session.

A second workflow turns on session learning: caliber learn install captures tool usage, failures, and your corrections, then distills the patterns into CALIBER_LEARNINGS.md — tagged [correction], [gotcha], [fix], [pattern], [env], or [convention] and deduplicated automatically.

Tips

• Generation uses the AI subscription or API key you already have — Claude Code or Cursor seats work, or bring an Anthropic, OpenAI, MiniMax, or Vertex AI key. Bootstrap and scoring never call an LLM.
• Telemetry (command names and durations, no code) goes through PostHog; set CALIBER_TELEMETRY_DISABLED=1 to opt out.
• Pre-commit auto-staging can be disabled for signed-commit or minimal-diff workflows: git config caliber.autostage false.
• caliber uninstall removes everything Caliber added — hooks, generated sections, skills, learnings — while preserving your own content.

When Not to Use This

You need Node.js 20+, and the generation step needs an AI subscription or API key (bootstrap and scoring do not). If you maintain a single hand-curated context file for a single agent and review it every release, the multi-platform sync machinery is more than the problem calls for — the deterministic score alone may be the useful part.

See the leaderboard for more skills.