# brainstorm — a skill that has to end in a decision A divergence → evaluation → bias-audit → **decision** pipeline for coding agents (Claude Code, or anything that loads a skill/instruction file). Download page: --- ## What's in here ``` SKILL.md the skill definition — this is the file the agent loads scripts/verify_scores.py all the arithmetic; the enforcement gate lives here references/criteria-sets.md axis definitions and scoring anchors example/demo_state.json a filled-in run you can execute immediately ``` ## Try it in 60 seconds ```bash python scripts/verify_scores.py example/demo_state.json ``` You will get a ranking with a `PRELIMINARY` watermark on it. That is the point — the bias audit has not run. Now try to clear it dishonestly: ```bash python scripts/verify_scores.py example/demo_state.json --bias-done ``` It refuses, with exit code 2, because `bias_audit.findings` is empty. Write findings into the state file, run it again, and the watermark clears. Start your own run with: ```bash python scripts/verify_scores.py my_run_state.json --init --set TECH ``` ## Why it is built this way An earlier, looser version of this process ran 9 times. It produced **604 ideas** and converted **zero** of them into a chosen next step. Its bias-audit phase was declared mandatory in the instructions and actually ran on **5 of 9** runs. Both numbers come from the same cause: the enforcement lived in prose, and prose is a suggestion. The two mechanisms that fixed it are the only genuinely novel things here: 1. **A watermark the model cannot clear by asserting.** Rankings print as `PRELIMINARY` until a script records the audit — and the script refuses to record an audit with no findings attached. 2. **A forced pick at the end of every run.** "Pick 3–5 now, or tell me *when*." Deferral stays allowed; it just stops being silent. Everything else — the six axes, the weight profiles, the wildness quota — is ordinary brainstorming structure. Those two are what changed the outcome. ## Requirements Python 3.7+. No dependencies. ## Provenance Drafted by an AI agent that runs a real company, reviewed line by line by the human who runs it. The numbers above are from that system's own logs. This is a port: internal paths, registers and company references have been removed, and the axis vocabulary has been generalised. The enforcement logic is unchanged. Longer write-up: ## License MIT.