Your AI agent is lying to you — and your CLAUDE.md is why

--- cover_image: https://raw.githubusercontent.com/rely-ai-org/caliber/master/assets/social-preview.png --- > **TL;DR** — Your AI coding agent's quality is capped by the quality of its context. Most devs have stale, generic, or missing CLAUDE.md / .cursorrules files. [Caliber](https://github.com/rely-ai-org/caliber) fixes this in one command — scans your repo, generates tailored configs, recommends MCPs, and gives you a 0-100 setup score.<br>---<br>## The problem nobody talks about<br>Everyone's debating **which AI coding tool is best** — Claude Code vs Cursor vs Codex. Meanwhile, the real reason most developers aren't getting great results is upstream of all of them:<br>**Their project context is wrong.**<br>Your `CLAUDE.md` was written in 20 minutes when you first set up the project. Your `.cursorrules` was copy-pasted from a Reddit thread. Neither has been touched since.<br>And your AI agent is making decisions based on that stale, inaccurate information every single session.<br>## What bad context actually looks like<br>Here are the most common failure modes I've seen:<br>**🚨 Stale architecture** — Your `CLAUDE.md` says you use REST APIs but you migrated to GraphQL 3 months ago. The agent keeps generating REST patterns.<br>**🚨 Contradictory rules** — Old rules say "use CommonJS", newer ones say "use ESM". The agent picks one arbitrarily.<br>**🚨 No MCP coverage** — You're running PostgreSQL and there's a great Postgres MCP that would let your agent query your schema directly. You've never heard of it.<br>**🚨 Config drift** — You refactor every week. Your AI config was updated once, on day one.<br>**🚨 Team inconsistency** — One dev has MCPs set up, another doesn't. Rules differ across machines. There's no source of truth.<br>## Meet Caliber<br>[**Caliber**](https://github.com/rely-ai-org/caliber) is a CLI tool that solves this with one command. It scans your project and:<br>- ✅ Generates a tailored `CLAUDE.md` with your actual stack, architecture, and commands<br>- ✅ Creates `.cursorrules` / `.cursor/rules/*.mdc` matching your dependencies<br>- ✅ Recommends MCPs you should install based on what you're running<br>- ✅ Deletes stale rules that contradict your current code<br>- ✅ Scores your setup **0–100** across 6 dimensions<br>And it can run continuously — so configs stay fresh as your code evolves.<br>## See it in action<br>Here's what `caliber init` looks like on a real Next.js + TypeScript project:<br>```bash<br>$ caliber init<br>Scanning project structure...<br>Detected: TypeScript, React, Next.js, Tailwind CSS<br>Detected: 847 files, 12 dependencies with AI relevance<br>Config files:<br>+ create CLAUDE.md project context<br>+ create .cursorrules cursor rules<br>~ modify .cursor/rules/testing.mdc outdated patterns<br>- delete .claude/rules/old-api.md stale, contradicts code<br>Skills:<br>+ create .claude/skills/deploy.md deploy flow<br>+ create .cursor/skills/review.md code review<br>MCP Recommendations:<br> • @modelcontextprotocol/server-postgres (detected: pg in package.json)<br> • @modelcontextprotocol/server-github (detected: .git, gh-cli)<br> • @upstash/context7-mcp (detected: React 18+)<br>```<br>Notice it's **not just adding** files — it also deletes stale ones and modifies outdated patterns.<br>## The setup score: caliber score<br>This is my favorite feature. Run `caliber score` on any project and get a deterministic 0–100 grade — **no LLM needed, works offline**.<br>```bash<br>$ caliber score<br>Score: 87/100 Grade: A<br>Existence 23/25 • Config files present, cross-platform parity<br>Quality 22/25 • Commands documented, no bloat, no vague rules<br>Coverage 18/20 • Dependencies & services reflected in configs<br>Accuracy 13/15 • Documented commands & paths actually exist<br>Freshness 8/10 • Config recency, no leaked secrets<br>Bonus 3/5 • Hooks, AGENTS.md, OpenSkills format<br>```<br>This gives you an **objective baseline** before onboarding a new dev, switching AI tools, or auditing your setup.<br>## The full CLI — 4 commands, that's it<br>| Command | What it does |<br>|---|---|<br>| `caliber init` | Scan repo, generate/update all config files |<br>| `caliber score` | Rate your setup 0–100 (offline, no LLM) |<br>| `caliber recommend` | Discover MCPs and skills for your stack |<br>| `caliber config` | Set provider, API key, model |<br>**Works with:** Claude Code, Cursor, OpenAI Codex<br>**No API key needed** with Claude Code or Cursor — uses your existing subscription.<br>## Why this matters for teams<br>This isn't just a solo-dev tool. The consistency problem is **worse at the team level**:<br>> *"One dev has MCPs configured, another doesn't. Cursor rules differ across machines. Nobody knows which CLAUDE.md is the canonical one."*<br>With Caliber, you commit your configs to git like any other file. Every developer who runs `caliber init` gets the same baseline — and the same AI agent experience. New team members are set up in 30 seconds.<br>## Get started in 30 seconds<br>```bash<br>npm install -g @rely-ai/caliber<br>caliber init<br>```<br>That's it. No API key needed if you're on Claude Code or Cursor.<br>🔗 **GitHub:** https://github.com/rely-ai-org/caliber<br>📦 **npm:** https://www.npmjs.com/package/@rely-ai/caliber<br>💬 **Discord:** https://discord.gg/XUNaJEsw<br>MIT licensed. Open source. Your code never leaves your machine.<br>---<br>*Built at [Rely AI](https://caliber-ai.up.railway.app/). If this saved you time, a ⭐ on GitHub goes a long way — and PRs are very welcome.*