Context Autopilot reads your past Claude sessions — everything you repeated, corrected, or rejected — plus the notes Claude already keeps about your projects, and turns them into rules every future chat starts with. You approve every rule before anything is written.
You've spent months teaching Claude how you like things done — one chat at a time. Autopilot makes those lessons permanent.
Your chat history, your CLAUDE.md files, and the memory Claude keeps for each project. The lessons are already sitting in there — Autopilot finds them.
Approved rules go into the project's rules file and your personal global one — so every new session, in every project, starts already knowing how you work. You stop re-explaining yourself.
Watching and scanning use zero AI — it's plain file-reading on your own computer. AI runs once, only at the moment rules are written, through the Claude plan you already have. No API key, no separate bill.
A silent check runs when a session starts. Only when enough new patterns have built up does Claude offer to save them — at a natural pause, not mid-task. And nothing is ever written without your yes.
Once installed, Claude picks the right tool from your plain words. Things people actually say:
Claude answers with suggested rules and the exact words you once said that justify each one. You reply "accept 1 and 3" — done.
Your chat history is a literal record of what Claude got wrong and what you had to say to fix it. Autopilot turns that record into rules Claude keeps.
On your own computer, it reads your past sessions and saved project notes, looking for three things: instructions you gave more than once, moments you corrected Claude, and actions you refused. Nothing is uploaded anywhere.
Only the repeated patterns — not your whole history — get turned into plainly-worded rules. Each one comes with the exact words you once said, quoted as proof. Generic advice gets filtered out.
Accepted rules are written into a clearly marked section of your rules files. Anything you wrote by hand is never touched, rejected rules are never suggested again, and nothing is written without your yes.
Pointed at a real project's chat history, Autopilot surfaced rules like these — each one something the person had already taught Claude the hard way:
Research on machine-generated CLAUDE.md files found they reduce task success and run up your token bill — automatic scans produce generic filler Claude then has to wade through. A focused 50-line file beats a sprawling 1,000-line one.
Every time you told Claude "no — like this," you wrote down exactly what you want. Autopilot only proposes rules your own words support, quotes them as proof, and prefers a few strong rules over a pile of weak ones.
Rules land in both CLAUDE.md and AGENTS.md, so Claude Code, Cursor, Copilot, Codex, and 30+ other agents all benefit. It also catches the reverse problem: rules your project has outgrown — references to files that no longer exist — before they mislead an agent.
It looks across all your projects for rules about how you personally work — "explain in plain English", "don't build while I'm brainstorming" — and keeps them in your one global rules file, so every project benefits. It also reads the memory files Claude keeps per project and lifts the universal lessons up to that global file. No other tool does evidence-based personal context.
An independent, curated map of the context layer — who does what, without the vendor slant. Updated continuously.