Stage Gate Loops — Development skill for Claude Code

stage gate loops

**Every stage of an agent's work ends at a gate. The gate passes, the next stage runs. The gate fails, the line stops.**

[](https://github.com/heyytars/stage-gate-loops/actions/workflows/ci.yml) [](LICENSE) [](pyproject.toml) [](action.yml)

`sgl` is a small runner for that pattern. A gate is any command: exit `0` means pass, anything else stops the run. No model grades the work, so there is nothing to argue with. A schema either validates or it doesn't. Tests either pass or they don't.

It works with any agent: Claude Code, Codex, your own scripts, or plain cron. PyYAML is the only dependency.

Try it in a minute

pip install git+https://github.com/heyytars/stage-gate-loops
git clone https://github.com/heyytars/stage-gate-loops
cd stage-gate-loops/examples/blog-publish && sgl run pipeline.yaml

The draft broke one rule. The gate caught it at stage 3, the exact failing line went back to the agent, and only stage 3 ran again. Research did not repeat, and publish never saw a bad draft.

The examples use offline stub agents, so they run free and give the same result every time. Point them at a real model with `AGENT_CMD="claude