Known-bad fixtures
The independent-gate judge compares the candidate SKILL.md to each fixture.
Match means the same shape, not byte identity and not the same name:.
Any match → Verdict: FAIL. Do not invent extra shapes. Do not skip a
fixture because the candidate added sentences.
Fixture: stub-happy
---
name: improve-skill
description: Improve agent skills.
---
# Improve Skill
Review the target skill and make it better. Be thorough. Consider clarity
and completeness. Commit your changes when you are happy with them.Match if: frontmatter plus an intent paragraph; no ordered cycle (select / baseline / diagnose / one-change / independent gate / commit-on-pass); description has no WHEN; the author decides it is better ("when you are happy").
Fixture: numbered-gate-free
---
name: improve-skill
description: Improve a SKILL.md. Use when asked to improve a skill.
---
# Improve Skill
1. Read the skill.
2. Think about improvements.
3. Edit it.
4. Commit.Match if: numbered steps with no off-tree baseline, no diagnosis lens, no one-change limit, and no independent FAIL path; commit is the last step.
Other shapes (no full fixture)
| Shape | Match if |
|---|---|
| triggerless | description has no WHEN and no quoted utterances |
| sibling-collision | load-steals duration work / code slop / graphs / metric loops without naming those siblings |
| placeholder | body is TODO or "write the procedure later" |
| ask-back | user must supply the diagnosis or the change |