Warmboot — AI skill for Claude Code

Warmboot Light

Warmboot Light gives AI coding agents memory and rules across sessions. Without shared state, a new session can start from zero: agents forget decisions, redo finished work, trust test results that are no longer valid and skip review. Warmboot Light keeps a short contract and the task's state as plain text in the repository, read through AGENTS.md or CLAUDE.md, with no service and no dependencies to install.

Designed and directed by Tiago Fonseca; implemented and reviewed with coding agents under this protocol. This is the compact edition of Warmboot; [Which edition?](#which-edition) compares it with [Warmboot Full](https://github.com/korreio/warmboot-full).

**Status: experimental.** Warmboot is a collection of practices for running coding agents. It makes no promise of better results. Use it at your own risk and take what is useful.

Mermaid source

flowchart TD
    accTitle: Agents lose context; Warmboot installs nine files; each session reads, works, proves and hands off
    P["Agents lose scope, overwrite work, reuse stale results"] --> I["init installs nine plain-text files"]
    I --> S
    subgraph S["Each session, then the next"]
        direction LR
        R["Read the task and the core"] --> W["Work inside the named scope"]
        W --> V["Prove: record what ran, on which inputs"]
        V --> H["Hand off: the next session checks those inputs"]
    end

Try it

Needs a POSIX-like platform, Python 3.11 or newer and an existing project directory given as an absolute path. Run from this checkout: preview (writes nothing), apply, then check:

python3 bin/warmboot init /absolute/path/to/project
py