Conciseness

The body is dense and high-signal — a terse sh loop block, a gate-rule table, hard rules — with no explanation of concepts Claude already knows. However, it inlines time-sensitive details ('always pinned (`npx -y human-review@0.6.0`)', 'See `audit/human-review-2026-08/`') outside any old-patterns/deprecated section, and rationale clauses like 'a mistyped severity silently downgrades to NIT, so a real blocker can be lost to a typo' could be trimmed. Anchor 4 (efficient, minor trimming possible) fits better than 5.

Actionability

Fully executable, copy-paste-ready commands with exact script paths and flags: 'python3 $S/human_gate.py open plan.md --launch', 'collect plan.md --output json', 'close plan.md', plus '--sample' covering the whole loop in ~1s. Exit-code semantics ('exit 2 = NOT done', '0 clear·2 blocked·3 collect·4 none') make outcomes unambiguous — matches anchor 5.

Workflow Clarity

The loop is a clear sequence (open → END TURN → status → collect → apply every item → close) with explicit validation checkpoints: the exit-2 gate, G4 (sidecar changed after last collect), G7 (integrity), and round-cap exhaustion that 'escalate[s], never pass[es]'. Batch feedback handling includes validation (quotes checked against raw and rendered text), so no cap applies; anchor 5's explicit validation and feedback loop are present.

Progressive Disclosure

References are one level deep, clearly signaled, and real: 'Reasoning lives in `references/` — human-in-the-loop canon, feedback batching, loop discipline' (all three files exist) and 'Worked example and JSON contract are in `assets/`'. Minor gap: the cited 'audit/human-review-2026-08/' path is absent from this bundle, and some inline rationale (the derivation note, forcing-question commentary) could live in references. Anchor 4 fits; not 5 because of the dangling pointer.

Specificity

The description lists multiple concrete actions — 'Builds a single-file HTML review page, collects batched feedback as a structured artifact… runs a gate that refuses to close while a BLOCKER is open, the reviewer is unnamed, or nobody has reviewed at all' — with comprehensive coverage and third-person voice. It matches the anchor-5 example's completeness; a 4 would require minor gaps in coverage, which are absent.

Completeness

Both 'what' (build review page, collect batched feedback artifact, run gate with named refusal conditions) and 'when' (explicit 'Use when a plan, spec, RFC… needs human sign-off before shipping, or on requests such as…' with concrete trigger phrases) are answered explicitly, plus NOT-for exclusions. This is the anchor-5 pattern exactly.

Trigger Term Quality

It includes natural phrases users would actually say, quoted directly: 'get sign-off', 'have someone check this', 'hold until reviewed', 'needs approval first', plus artifact-type keywords (plan, spec, RFC, report, landing page, migration). This matches anchor 5's comprehensive coverage including phrasing variations.

Distinctiveness Conflict Risk

It carves a clear niche (human sign-off/approval) and explicitly disambiguates the nearest conflicting skills: 'NOT for making AI text sound human (use content-humanizer or behuman). NOT for reviewing code diffs (use md-review or code-reviewer)'. Distinct triggers and minimal conflict risk match anchor 5.