Conciseness

Lean and information-dense; it assumes Claude's competence (no general explanations of NER, model cards, or equalized odds) and every section — intro, metrics, quick start, workflow, hand-offs, gotchas — earns its place.

Actionability

Quick start provides copy-paste-ready, fully executable code (import, fairness_report call with concrete args, result printing, and an assert gate) covering the common audit case.

Workflow Clarity

A clear six-step sequence with analytical checkpoints (read leakage first, compute disparity, locate worst_group, document the gap), an explicit assert verification gate, and a gotchas section that serves as a checklist; the operation is read-only so the destructive-cap does not apply.

Progressive Disclosure

Well-organized into clearly headed sections with one-level-deep, clearly signaled skill hand-offs (building-gold-corpus, evaluating-with-leakage-gates, authoring-model-cards, benchmarking-clinical-ner) and a source-of-truth pointer; no nested references and no bundle files needed.

Specificity

Names the domain (OpenMed NER/de-identification) and multiple concrete actions — auditing performance disparities across subgroups, surfacing documentation gaps, producing equalized-odds disparity numbers — via a named tool (openmed.eval.fairness_report); third-person voice.

Completeness

Explicitly answers both what (audit subgroup performance disparities) and when (a multi-clause "Use when..." plus an explicit "Trigger on..." list with concrete trigger phrases).

Trigger Term Quality

Comprehensive natural trigger list including synonyms and a domain standard: "fairness", "subgroup", "bias audit", "disparity", "equalized odds", "under-protected group", "per-group recall", "STANDING Together".

Distinctiveness Conflict Risk

A clearly bounded niche — OpenMed fairness/de-id under-protection audits — with distinctive triggers, making conflict with unrelated skills unlikely.