quality loops

This skill implements structured iteration loops for high-stakes outputs where a model's first draft is insufficient. It enforces a strict separation between the generator and the critic, ensuring work is not shipped until it clears a measurable, numeric quality bar. By using defined patterns like Multi-Critic and Adversarial Critique, it prevents circular arguments, hallmarked errors, and superficial "looks good to me" approvals.

What it does

• Enforced Scoring requires every iteration to have a numeric progress metric rather than vague prose feedback.
• Generator-Critic Separation ensures the model that writes the content is not the same instance that approves it.
• Multi-Dimension Auditing tracks independent scores for correctness, style, and safety without letting one high score mask a failure in another.
• Adversarial Stress-Testing identifies edge cases and security exploits by actively attacking the draft until no unrebutted objections remain.
• Stall Detection monitors for repeated failed rewrites, triggering escalation rather than infinite, unproductive loops.

How it works

1. Pattern Selection Choose between Generate-Critique-Rewrite, Score-and-Retry, Multi-Critic, Adversarial Critique, or Judge Ensemble based on the complexity of the task.
2. Loop Initialization Define a specific numeric target value and a maximum iteration budget using the loop-ledger tool.
3. Critic Pass Run independent critics or automated scorers to identify specific issues and generate a progress score.
4. Targeted Rewriting The generator iterates on the draft, specifically addressing the issues list provided by the critic.
5. Gate Exit The process terminates only when the numeric threshold is cleared, the budget is exhausted, or the critique-gate plugin halts the session.

Frameworks & tools

Built for use with loop-ledger and the critique-gate plugin. It utilizes standard GitHub-flavored Markdown for reporting and works across LLM providers supporting tool-calling and system prompt overrides.

Why this beats prompting it yourself

Generic "critique this" prompts often result in polite, superficial feedback that fails to improve the actual content. This skill forces numeric accountability and prevents the model from closing a task before the quality bar is demonstrably met.

Use cases

• Technical Postmortems refining incident reports until they meet clarity and completeness scores.
• Security Audits running adversarial loops to find exploits in design documents or code.
• PR Reviews scoring code changes independently for style, logic, and performance before merging.
• Legal/Policy Drafting ensuring documents survive a judge ensemble where multiple critics must reach consensus.

Known limitations

Requires a numeric metric to function; it cannot iterate on purely subjective tasks that lack a scorable dimension. It is not intended for exploring multiple design variations.