coldstartlisted
Install: claude install-skill thomasog/coldstart
# Coldstart
You run this once, at the start of a project. You leave behind a folder that any agent can pick up on any machine, arriving cold: everything that matters lives inside the folder, the instructions follow the open AGENTS.md standard, and the AI's working files stay out of the user's way.
This session's product is a tidy folder, not progress on the project. The project's work starts in the next session. AGENTS.md holds what the next agent cannot recover from the files: the user's goal, who is involved, and the dates and rules the user states. Every file must earn its place, so write little and say specific things. The reasoning behind each piece is in `references/rationale.md`; read it if you need to justify or adapt a choice.
## What you create
```
<project>/
├── AGENTS.md # the only new file at the root; any agent reads it
├── .ai/
│ ├── HANDOFF.md # where we stopped and what comes next
│ └── memory/
│ └── MEMORY.md # durable facts
├── .gitignore # only if the folder is or will be in git
└── .claude/ # only if the user uses Claude Code
├── CLAUDE.md # imports AGENTS.md, the handoff and the memory
└── settings.json # plans in .ai/; auto memory off, so earlier auto memories stop loading; .env blocked
```
Create nothing else. `.ai/plans/`, `.ai/tools/`, `.ai/archive/` and `.ai/scratch/` come into being the first time they're needed. Codex, Cursor, Copilot and Antigravity read AGEN