i2r-skill-quality-mode

Purpose

This mode governs the quality loop for I2R itself — not the requirements it produces. It encodes skill-TDD: before trusting any i2r-* skill inside an agent, you must observe how a naive agent behaves without it, build evals that catch the failure, then verify the skill closes the gap. Sources: Anthropic skill-creator pattern; Superpowers writing-skills eval discipline.

Vendor-not-install: patterns are extracted from those sources and implemented as local i2r-* logic. No runtime calls to external skills.

Core principle: skill-TDD

1. Observe naive failure first. Run the target task without the candidate skill. Record what goes wrong (implementation leak, vague NFR, etc.).
2. Write a failing eval. Add a pressure scenario under evals/pressure-scenarios/ that reproduces the failure.
3. Author/fix the skill. Update the SKILL.md or its references/.
4. Verify the eval passes. Run i2r.py evals.run — all stages of the pipeline must turn green before the skill ships.
5. Gate on description-trigger accuracy. Run evals/trigger-tests/ to confirm the skill fires on the right prompts and not on near-misses.