Retro Speedlab — Development skill for Claude Code

Retro Speedlab Cookiecutter

[](https://github.com/datenwissenschaften/retro-speedlab/actions/workflows/ci.yml)  [](LICENSE)

A Cookiecutter template that scaffolds a game package for [Retro Speedlab](https://www.retrospeedlab.com), the lab where the Laya decision model learns to beat retro games by itself. The generated package starts as a skeleton that knows nothing about its game. Scheduled Claude Code lab runs grow it from what they verify in the emulator, and the [Retro Speedlab Core](https://github.com/datenwissenschaften/retro-speedlab-core) engine turns it into a learning agent.

What it generates

flowchart TD
    CC[Cookiecutter] --> GEN[Game package skeleton]
    GEN --> A[Plain NES buttons]
    GEN --> B[Empty GameRam]
    GEN --> C[One Play state, no reward]
    GEN --> D[Configuration + tests]
    GEN --> E[Daily lab run + beat-nes-game skill]
    GEN --> F[Dockerfile + Dokku deployment]
    E -. grows .-> B
    E -. grows .-> C

The template deliberately ships no game knowledge. Everything a game needs (RAM map, phases, detectors, hints, rewards, savestates for new levels) is built by the lab runs, which follow the bundled `.claude/skills/beat-nes-game/SKILL.md`.

Quick start

pipx install cookiecutter
cookiecutter https://github.com/datenwissenschaften/retro-speedlab
cd retro-speedlab-game
poetry install
poetry run python app.py

Put the ROM Stable Retro expects for the chosen game into `roms/` first; it is imported automatically when training starts.

Template variables

Variable	Purpose
project_name	Human-readable project name
project_slug	Distribution and directory name, derived from project_name
game	Stable Retro game id, for example