datenwissenschaften/retro-speedlab
Cookiecutter template for Retro Speedlab game packages: a skeleton that scheduled Claude Code lab runs grow until the TAS search beats the game.
stack: Python
ARCHETYPE
YOLO Cowboy
Permissions? Never heard of 'em. Ships at the speed of `--dangerously-skip-permissions`.
Copy this rig
# review before running: this installs third-party code
$ npx degit datenwissenschaften/retro-speedlab/{{cookiecutter.project_slug}}/.claude ./rig-retro-speedlab # inspect, then merge into .claude/ MCP servers are added to Claude Code at local scope; env vars are shown as YOUR_… placeholders — we never store values. Files are fetched with degit into a separate folder so you can review before merging.
$ npx degit datenwissenschaften/retro-speedlab/{{cookiecutter.project_slug}}/.claude/skills/beat-nes-game .claude/skills/beat-nes-game $ curl -fsSL --create-dirs -o .claude/agents/engineer.md https://raw.githubusercontent.com/datenwissenschaften/retro-speedlab/main/%7B%7Bcookiecutter.project_slug%7D%7D/.claude/agents/engineer.md $ curl -fsSL --create-dirs -o .claude/agents/researcher.md https://raw.githubusercontent.com/datenwissenschaften/retro-speedlab/main/%7B%7Bcookiecutter.project_slug%7D%7D/.claude/agents/researcher.md $ curl -fsSL --create-dirs -o .claude/agents/scout.md https://raw.githubusercontent.com/datenwissenschaften/retro-speedlab/main/%7B%7Bcookiecutter.project_slug%7D%7D/.claude/agents/scout.md $ curl -fsSL --create-dirs -o .claude/agents/verifier.md https://raw.githubusercontent.com/datenwissenschaften/retro-speedlab/main/%7B%7Bcookiecutter.project_slug%7D%7D/.claude/agents/verifier.md
Merge into .claude/settings.json (project) or ~/.claude/settings.json (user). Hook commands reference scripts in the source repo — copy those too.
{
"permissions": {
"deny": [
"Bash(sudo:*)",
"Bash(git push:*)",
"Bash(rm -rf /app:*)",
"Edit(/app/working/**)",
"Write(/app/working/**)"
]
}
} Skills (1)
Subagents (4)
| engineer model: inherit | Writes the smallest change and its test. Use once a finding is measured, to change states, rewards, dynamics() or RAM fields in the game package and pin the change with a test; returns the diff summar |
| researcher model: inherit | Checks published routes and RAM maps. Use to learn how the top human speedruns route a level, which skips the category allows, and which RAM addresses published maps name; returns findings with source |
| scout model: inherit | Measures RAM and behaviour in the emulator. Use to verify a RAM address, a transition or a mechanic with a small probe from a savestate reached from power-on; returns measured values with the committe |
| verifier model: inherit | Runs the tests and the power-on replay. Use after every commit the lead wants to ship, to run agent/verify.sh (ruff, pytest, the movie's verify) and review the diff for leanness; returns pass or fail |
Permissions
deny (5)
Bash(sudo:*)
Bash(git push:*)
Bash(rm -rf /app:*)
Edit(/app/working/**)
Write(/app/working/**)
ask (0)
—
allow (6)
Bash
Read
Edit
Write
Glob
Grep
Similar rigs
leavemagic-cyber/earned-confidence
A behavior contract for AI coding agents — every rule added after a real failure. Ships with a verifier agent that cannot edit what it judges.
Pragmatist 3.6k tok · · Q5
doublegate/Master-Claude
Curated, project-agnostic synthesis of AI-agent knowledge across a multi-project workspace, plus POSIX-sh tooling to install and keep it in sync inside any project. One source of truth for Claude Code, Codex, and Gemini (AGENTS.md + symlink
Pragmatist 985 tok · · Q0
ivan-amiqa/claude-code-agent-orchestration
Five Claude Code sub-agents: one boss, cheaper specialists, and a script to measure what it saves you.
Orchestrator 396 tok · · Q30
j0yen/dotfiles
The claude-self cluster: machine-personal CLAUDE_SELF.md, its SessionStart hook, and the claude-self CLI.
YOLO Cowboy 6.6k tok · · Q10