Context
Absorb candidate recorded in PR #870 (which shipped claude-config:audit-instructions and superseded audit-model-fit, issue #800). The superseded skill's evals carried a dedicated step-list-culled-not-preserved case (its C2 check): a 9-step numbered PR procedure must be flagged over-prescriptive and culled to intent + hard constraints (real ordering, safety gates, external contracts kept), not preserved verbatim. audit-instructions covers over-prescriptive step lists under check I8 (Model-era re-audit) but its evals/evals.json (5 cases) has no case exercising step-list culling specifically — the nuance "cull to intent + hard constraints, keep genuinely load-bearing steps" is untested.
Proposed work
- Add a step-list-culling eval case to
plugins/claude-config/skills/audit-instructions/evals/evals.json, modeled on the superseded case but phrased against the I8 catalog: identifies the step list as an I8 finding, recommends culling to intent + hard constraints, keeps steps encoding real ordering/safety/contract.
- Validate against the evals schema (
skill-quality:check validate-evals path); patch-bump claude-config + CHANGELOG per changelog-parity.
Acceptance criteria
References
Metadata
| Field |
Value |
| Category |
unspecified |
| Area |
unspecified |
| Ecosystem |
unspecified |
Context
Absorb candidate recorded in PR #870 (which shipped
claude-config:audit-instructionsand supersededaudit-model-fit, issue #800). The superseded skill's evals carried a dedicatedstep-list-culled-not-preservedcase (its C2 check): a 9-step numbered PR procedure must be flagged over-prescriptive and culled to intent + hard constraints (real ordering, safety gates, external contracts kept), not preserved verbatim.audit-instructionscovers over-prescriptive step lists under check I8 (Model-era re-audit) but itsevals/evals.json(5 cases) has no case exercising step-list culling specifically — the nuance "cull to intent + hard constraints, keep genuinely load-bearing steps" is untested.Proposed work
plugins/claude-config/skills/audit-instructions/evals/evals.json, modeled on the superseded case but phrased against the I8 catalog: identifies the step list as an I8 finding, recommends culling to intent + hard constraints, keeps steps encoding real ordering/safety/contract.skill-quality:check validate-evalspath); patch-bump claude-config + CHANGELOG per changelog-parity.Acceptance criteria
evals/evals.jsongains a step-list-culling case with expectations covering: I8 identification, cull-to-intent recommendation, load-bearing steps preserved.References
git show 9533a14^^:plugins/claude-config/skills/audit-model-fit/evals/evals.json(casestep-list-culled-not-preserved, id 7)plugins/claude-config/skills/audit-instructions/reference/criteria.md(check I8)Metadata