Conciseness

The body is dense with non-inferable specifics (UnavailableReason cases, OS27 error-migration details, evaluation-framework wiring) and avoids teaching basics, but the OS 26/27 migration story is told roughly three times ('Search BOTH error spellings...', Pattern 4, Pattern 14) and Patterns 4 and 5 repeat the same generic-catch grep. Minor instances of repetition that could be trimmed — anchor 4, not 5.

Actionability

Every pattern ships literal, copy-paste-ready grep regexes ('import\s+FoundationModels', 'SystemLanguageModel\.default\.availability'), complete executable Swift fix snippets ('guard SystemLanguageModel.default.availability == .available else {...}', '.disabled(session.isResponding)'), and specific per-match verification instructions. Matches anchor 5 ('fully executable; copy-paste ready; covers common cases').

Workflow Clarity

Five clearly sequenced phases with explicit validation checkpoints: 'Present this map in the output before proceeding', 'For every grep match, use Read to verify the surrounding context before reporting', plus Output Limits and a False Positives section forming a review checklist. A built-in feedback loop (grep recall → Read verify) matches anchor 5; the audit is read-only, so no destructive-operation cap applies.

Progressive Disclosure

References are clearly signaled and one level deep ('See `axiom-ai (skills/foundation-models-ref.md)` for the full migration table'), and the migration table itself is appropriately deferred rather than inlined. However, no bundle files exist in this skill, all 14 detection patterns are inlined in a ~485-line monolith, and the referenced paths live in a different bundle ('axiom-ai') so their existence cannot be verified here — anchor 4 ('good structure; references mostly clear; minor organization gaps'), not 5.

Specificity

Names the domain ('Foundation Models review', 'on-device AI audit') and several concrete detection targets ('missing availability checks, manual JSON parsing, and guardrail handling'), but coverage is not comprehensive — session lifecycle, concurrency, and evaluation-suite detection from the body are absent. Anchor 4 ('lists several specific actions; minor gaps') fits; not 5 because the action list is incomplete, not 3 because more than 1-2 concrete actions are given.

Completeness

The 'when' is explicit and concrete ('Use when the user mentions Foundation Models review, on-device AI audit, LanguageModelSession issues...') and the 'what' is present via the 'including missing availability checks, manual JSON parsing, and guardrail handling' list, but there is no direct leading capability statement (e.g., 'Audits Foundation Models code for...'). Fits anchor 4 ('both what and when; one could be more explicit') rather than 5.

Trigger Term Quality

Includes natural user phrasing plus specific API keywords: 'Foundation Models review', 'on-device AI audit', 'LanguageModelSession issues', '@Generable/guardrail problems'. A few common natural terms are missing — notably 'Apple Intelligence' (used heavily in the body) and 'Apple' platform qualifiers — so it fits anchor 4, not 5.

Distinctiveness Conflict Risk

Apple-specific API names ('LanguageModelSession', '@Generable', 'guardrail', 'Foundation Models') carve out a clear niche with distinct triggers and minimal overlap with other skills, matching anchor 5.