Conciseness

The body is dense with non-obvious version-specific facts (e.g., "runtime obsoleted in 27.0", "Apple's own sample judge starts at Cohen's kappa −0.037") and pads nothing with concepts Claude already knows, but routing decisions are restated across four sections (AI Approach Triage, Decision Tree, Example Invocations, Critical Patterns) and the Opus 4.7 migration warning appears three times. Above 2 because the redundancy is structural repetition rather than explanatory filler; below 4 because the duplication exceeds minor instances.

Actionability

As a router skill the guidance is fully deterministic: exact file targets per situation ("Read: `skills/foundation-models-diag.md` (Pattern 5a...)"), exact symbols ("compatibleAdapterNotFound", "`provider(from:in:)`"), and explicit gates ("only after documented rung-1-4 failures"), with Example Invocations mapping natural user phrasings to files. The scoring-note allowance for instruction-only skills applies, and the guidance has no gaps for its purpose.

Workflow Clarity

A clear triage-first sequence ("First, determine which kind of AI the developer needs"), a numbered 14-step Decision Tree, and gating rules for the adapter path provide a well-sequenced workflow. Below 5 because there is no validation/feedback checkpoint on the routing decision itself, and the four parallel routing sections leave mild ambiguity about the canonical entry point.

Progressive Disclosure

SKILL.md acts as a router with ~20 one-level-deep `skills/*.md` references, each signaled with a concise content summary — good disclosure structure. Below 5 because none of the referenced files exist in the provided bundle to verify, and sections such as the adapter error-case lists inline detail that the referenced diag files presumably own.

Specificity

"implementing, testing, or evaluating" lists three concrete actions plus a specific technology list (Foundation Models, @Generable, LanguageModelSession, Tool protocol, eval suites, model-as-judge scoring, SpeechTranscriber, CoreML). Below 5 because the action list is narrow and omits debugging/diagnostics, a major use case the body routes heavily; above 3 because coverage extends well beyond 1-2 actions.

Completeness

Explicitly answers when ("Use when implementing, testing, or evaluating ANY Apple Intelligence, on-device AI, or speech-to-text feature") and what ("Covers Foundation Models, @Generable, LanguageModelSession, Tool protocol, eval suites, model-as-judge scoring, SpeechTranscriber, CoreML") with concrete trigger phrases, matching the top anchor. Not 4 because the 'when' clause is already maximally explicit and specific.

Trigger Term Quality

"Apple Intelligence", "on-device AI", "speech-to-text", and "SpeechTranscriber" are terms developers naturally say when they need this skill. Below 5 because common synonyms like "transcription", "dictation", and "SFSpeechRecognizer" are missing; above 3 because keyword coverage is good, not just partial.

Distinctiveness Conflict Risk

A clear Apple-specific niche with distinct triggers, but "on-device AI" and "CoreML" overlap with the custom-ML/CoreML skills that the body itself routes elsewhere, creating minor wrong-skill trigger risk. Below 5 for that overlap; above 3 because the niche is unmistakably Apple Intelligence/Foundation Models.