Skills, passes, and creative iteration

A repeatable iteration system:

• Skills (project instructions) that tell agents how your repo or workflow behaves.
• Passes: brainstorm → outline → draft → critique → revise, each with a checklist.
• Measurable improvement between passes (shorter, clearer, more consistent—not just “different”).

• Large skills ecosystem in the project culture—agents stay grounded when instructions live in-repo.
• AGENTS.md-style conventions (where you use them) align teams and models on rules.
• Multi-model setups let you route cheap models for breadth and stronger models for depth.

1. Write the skill as if onboarding a careful contractor: scope, inputs, outputs, pitfalls.
2. One pass per request until the workflow feels stable.
3. Critique with a separate prompt or model pass to reduce self-congratulation bias.
4. Version skills when behavior changes.

• A place for skills in your project layout (follow project docs).
• Criteria for “done” on each pass (length, tone, tests, human sign-off).

1. Define quality bars before generating volume.
2. Implement pass A; collect failures.
3. Adjust skill text from failures—not from vibes.
4. Add pass B; compare against pass A with the same inputs.
5. Freeze when improvements plateau.

• “Act as a harsh editor; do not rewrite—list issues by severity.”
• “Propose three variants with tradeoffs; recommend one.”
• “What skill text would have prevented this mistake?”

• Skills are short enough that people actually read them.
• Passes have clear entry/exit criteria.
• You can reproduce a prior run with the same inputs and skill version.

• Mega-skills that try to cover everything—split them.
• Skipping critique passes—same blind spots forever.
• Changing skills mid-flight without versioning—debugging becomes impossible.