Open source · Agent Skills
Clean Skills
Agent Skills with an executable Definition of Done.
The strongest reasoning goes into the contract, cheaper models do the work, and a fail-closed gate decides what is done. Agreement between models is not evidence. The skills work in Claude Code, Codex and any host that reads the Agent Skills format.
Get started
cargo install --locked clean-skillsclean-skills install --host claude --applyclean-skills install --host codex --apply- v0.3.1
- Claude Code and Codex
- MIT licence
- Strong models(done)
- Contract(done)
- Qualified checks(done)
- Executor(done)
- Gatewaiting for the gate…
Set up
One crate, both agents.
Install the crate, then copy the skills into Claude Code, Codex or both. Leave off the apply flag to see what install would write first. The skills’ helper and verification kit need Python 3.9+.
Three skills
clean-api-docs
Finds the library version your project actually locks or installs, reads that code and the official docs for that version, stops early, and labels where each fact came from.
clean-debate
Several strong models work out a task independently and correct each other. The result is a qualified contract that a capable, not necessarily the strongest, model can carry out.
clean-refactoring
Rebuilds a large application from first principles instead of refactoring it in place, against acceptance checks that accept a known-good result and reject a defective one.