What problem does it solve?
It prevents grade-inflation by forcing the AI to score its own output across multiple task-relevant quality dimensions before the result is shown to the user.
Core Features & Use Cases
- Five-dimension scoring (0-10): Evaluates output using the correct dimension set for code, design, documentation, specs/PRs/ADRs, or a default fallback.
- Worst-dimension gate: Uses the lowest sustained band as the truth signal, and treats any score below the regression threshold as a reason to fix and rescore.
- Pre-emit loop with evidence: Requires evidence-based justification per dimension (file paths, line numbers, concrete artifacts), then focuses iteration on the single lowest dimension.
Use cases include complex code changes, security-touching edits, substantial refactors, high-stakes documentation, and PR/ADR/release descriptions where mistakes are expensive post-emit.
Quick Start
Use the self-critique skill to score and verify the work you just produced, then emit only after all five dimensions meet the regression threshold.