Conciseness

All content is high-signal — no concept explanations, and the Anti-Rationalization table is dense non-obvious knowledge ("`Song.id` ... differs per device ... never use it as a cross-device key"). But routing information is triplicated: the Quick Reference table ("Camera capture, AVCaptureSession | See `skills/camera-capture.md`"), the dot Decision Tree ("what -> \"skills/camera-capture.md\" [label=\"camera capture\"]"), and the numbered Decision Tree ("1. Camera capture? → `skills/camera-capture.md`") all encode the same mapping — the dot graph is fully redundant with the numbered list. That is a whole section of redundancy, matching anchor 3 ("could be tightened") rather than anchor 4's "minor instances that could be trimmed"; not score 2 because nothing is padded or explains what Claude already knows.

Actionability

Directives are maximally concrete for an instruction-only router: "Read: `skills/camera-capture.md` (Pattern 8, deferred start)", "See `skills/carplay-templates-ref.md`" with named sections and pattern numbers, plus exact API-level fixes ("join your own session via `provider(from:in:)`, or `init(session:analyzerFormat:priority:)`"). Example Invocations ("User: 'My camera app takes a second before preview appears' → Read `skills/camera-capture.md` Pattern 8") cover the common cases. Per the rubric's code-vs-instruction note, absence of code is not penalized when guidance is this specific; not score 4 because no directive lacks the exact file/section/API to act on.

Workflow Clarity

The routing workflow is clearly sequenced: the numbered Decision Tree ("1. Camera capture? → ..."), Cross-Domain Routing resolves boundary cases explicitly ("Now Playing metadata/controls → **stay here** ... Background audio mode → **invoke axiom-integration**"), and Anti-Rationalization catches misdiagnoses. No validation checkpoints are required — routing is read-only, not destructive or batch. Score 4 rather than 5 because the auditor path is inconsistent: Example Invocations say "Launch: `camera-auditor` agent" while the inline note says "invoke the matching Codex skill ... Available: `axiom-audit-camera`" and step 14 says "→ `axiom-audit-camera`". Well above anchor 3 — the sequence is complete and unambiguous.

Progressive Disclosure

The body is a well-signaled, one-level-deep index: "See `skills/camera-capture.md` Patterns 8-9", "See `skills/carplay-templates-ref.md`", each mapped to a symptom row — textbook anchor 5 structure. But scored against the actual bundle: no `references/`, `scripts/`, or `assets/` directories exist, and none of the ~30 referenced `skills/*.md` paths resolve to a file in this bundle, so the references cannot be verified as real and the layout is nonstandard. Anchor 4 ("references mostly clear; minor organization gaps") is the best fit — not 5 because references cannot be confirmed against the bundle structure; not 3 because references are prominently signaled, not buried.

Specificity

"Use when working with camera, photos, audio, haptics, ShazamKit..." and "Covers AVCaptureSession, AVFoundation, Core Haptics, MediaPlayer..." name many domains and frameworks but state no concrete action verbs — the only verb is the generic "Covers". This matches anchor 2 ("Names the domain but actions are minimal or generic"): it is not score 3 because no 1-2 concrete actions (e.g. "set up capture sessions", "publish Now Playing metadata") are stated, and not score 1 because the language is concrete and domain-specific rather than vague or abstract.

Completeness

"Use when working with camera, photos, audio, ..." is an explicit, specific 'when' with concrete trigger phrases, and "Covers AVCaptureSession, AVFoundation, ..." answers 'what' — but only as a framework/domain list, not as actions. Anchor 4 ("Has both 'what' and 'when'; 'when' could be more explicit or specific") is the best fit in reverse: the 'what' could be more explicit. Not score 5 because the 'what' is a bare coverage list; not score 3 because the 'when' is explicit and detailed, not missing or weakly implied.

Trigger Term Quality

Rich natural keywords: "camera", "photos", "audio", "haptics", "screen recording", "casting to non-AirPlay routes", "SharePlay playback", "the user's Apple Music library", "lock-screen metadata", "CarPlay app design, templates, and navigation". Good-to-comprehensive coverage, but a few natural phrases users would say are missing ("video playback/export", "Now Playing", "Chromecast/Google Cast", "music playback"), so anchor 4 ("Good keyword coverage; a few natural terms missing") fits better than anchor 5, while clearly exceeding anchor 3's "missing common variations".

Distinctiveness Conflict Risk

Terms like "ShazamKit", "CarPlay", "SharePlay playback", "DockKit", and "casting to non-AirPlay routes" define a clear Apple-media niche with distinct triggers (anchor 5 territory), but "on-device music analysis" and "face grouping and video highlights" create minor overlap risk with sibling AI/vision skills, matching anchor 4 ("Mostly distinct; minor overlap risk with closely related skills"). Not score 3 — the niche is far more specific than "Works with document files".