Conciseness

There is significant duplication: the entire 'MANDATORY FIRST STEP' foundation check in Workflow step 0 is restated nearly verbatim as the first Rules bullet, and the internal-function exemption appears in both step 2 ('Never widen scope — an internal/admin function... stays internalQuery/internalMutation') and the Rules ('the one legitimate exception is an internalQuery/internalMutation'). The intro's defect statistics ('25 identity-from-arg + 13 missing-ownership-check + 6 PII-leak-by-argument = 44 of 214') add little operational value, and steps 0 and 2 are run-on paragraphs that could be tightened. Mostly purposeful content, but with clear tightening opportunities — anchor 3, not 4 (the redundancy is more than 'minor instances').

Actionability

Concrete guidance is strong: exact grep regex for shape (a), precise helper contracts ('throws 401 if ctx.auth.getUserIdentity() is null', 'throws 404 if doc is null, throws 403 if doc.ownerId !== identity.subject'), and a concrete verification command ('npx tsc --noEmit'). However, the canonical requireIdentity/requireOwner code itself is never shown — the skill says 'apply the canonical pattern from content/convex-expert.md verbatim — do not invent a new helper', yet no such file exists in the bundle, and shapes b-d get prose descriptions rather than the literal regexes promised by 'the 4 deterministic greps'. Mostly executable with a real gap — anchor 4, not 5.

Workflow Clarity

The sequence is explicit and well-gated: step 0 (foundation check with a clear branch: internalize-and-defer on foundationless apps), step 1 (deterministic scan), step 2 (harden), step 3 (VERIFY: 'run npx tsc --noEmit... Then re-run the step-1 scan to confirm 0 remaining hits'), step 4 (report with diffs). This is a batch-edit operation and it has both an explicit validation checkpoint and a feedback loop (re-scan until 0 hits), matching anchor 5's 'explicit validation steps; feedback loops for error recovery'.

Progressive Disclosure

There are no bundle files (no references/, scripts/, or assets/ directories), yet the body repeatedly directs the reader to 'content/convex-expert.md' and notes generation from 'content/capabilities/convex-authz.json' — paths that do not resolve inside the skill bundle, so the reference containing the canonical pattern to apply 'verbatim' is unnavigable. Meanwhile the operative content is inlined as two dense paragraph-walls (steps 0 and 2) that would be more usable split out. This matches anchor 3 ('references present but not clearly signaled; content that should be separate is inline') better than 4, where references would be mostly clear and resolvable.

Specificity

The description enumerates multiple concrete capabilities: 'identity-from-arg impersonation, missing per-document ownership checks, public queries leaking data by a client-supplied id, and writes into a parent/container the caller doesn't own', plus concrete actions ('Scans for the 4 shapes, applies requireIdentity/requireOwner, verifies with tsc'). This matches the anchor 'lists multiple specific concrete actions; comprehensive coverage' — not score 4, since there are no meaningful coverage gaps.

Completeness

Both 'what' (the four authz defect shapes and the scan/harden/verify pipeline) and 'when' (explicit 'TRIGGER on...' phrases plus a 'SKIP when there is no convex/ directory' boundary) are clearly and explicitly stated with concrete trigger phrases. This is exactly anchor 5; anchor 4's 'when could be more explicit' does not apply.

Trigger Term Quality

TRIGGER phrases 'secure my app', 'audit auth/authz', 'who can access this data' are natural user utterances, but common synonyms like 'permissions', 'authorization check', or 'lock down my app' are missing. Good coverage with a few natural terms absent — matches anchor 4 rather than 5's 'comprehensive coverage including synonyms'.

Distinctiveness Conflict Risk

The niche is narrow and well-fenced: Convex-specific authorization auditing, with an explicit SKIP ('no convex/ directory') preventing triggering on generic security-review requests. Distinct trigger phrases and platform-specific vocabulary (requireIdentity/requireOwner, convex/) give minimal conflict risk — matches anchor 5 rather than 4's 'minor overlap risk with closely related skills'.