Project context
One file is the brain of the project. Adapters are sticky notes for a specific agent. You are still in charge of anything that can blow up.
Agents have amnesia. Projects should not.
A coding agent is an AI that can read your folder, edit files, and run commands. Every new session, it only knows what you put in front of it. PROJECT.md is that briefing.
PROJECT.md
The project brain. Purpose, stack, features, safety rules, variable names. Every agent reads this. You edit this.
Adapters
Tiny files. “Read PROJECT.md. Wait for approval.” Claude wants CLAUDE.md. Codex wants AGENTS.md. Same brain, different nameplate.
Unzip. Read README.txt. Edit PROJECT.md.
PROJECT.md
You fill this in. Vision, stack, features, env names, and a hard list of things the agent may not do without a yes from you.
adapters/CLAUDE.md
Copy to CLAUDE.md in the project root if you are using Claude Code.
adapters/AGENTS.md
Copy to AGENTS.md in the project root if you are using Codex.
adapters/GROK.md + CURSOR.md + GEMINI.md
Same idea. Thin. They point back at PROJECT.md. They are not a second spec.
Or open PROJECT.md in the browser — also: the kit copy. Individual adapters are in the zip.
Names, never values
An environment variable is a labeled locked drawer. The label can go in PROJECT.md. The key cannot.
STRIPE_SECRET_KEY — Stripe server authentication credential
STRIPE_SECRET_KEY=sk_live_actual_secret_here
Plan. Then wait.
Read PROJECT.md. Do not write code yet. Write a numbered Phase 1 plan.
After each step include: what will be built, how I will test it, and any command.
Wait for my approval. Assume I am a beginner and explain terms in plain English.
Old Claude-only template
Still works in Claude Code. New projects should use the kit. Download the legacy CLAUDE.md template.