Conciseness

Mostly instruction-dense rather than explanatory, but it carries real padding: the ~20-line SCOPE LIMITS block is duplicated verbatim in both the Phase A and Round 2+ templates, the MCP polling instructions are repeated word-for-word, and the POSITIVE_THRESHOLD constant embeds editorial meta-commentary ("Earlier wording used 'or' + a stale verdict set; the AND form is authoritative"). It is not 2 because nearly all content is genuine operational guidance rather than concepts Claude already knows, and not 4 because the duplicated blocks are substantial removable tokens.

Actionability

Fully executable throughout: exact MCP tool names (mcp__gemini-review__review_start / review_status / review_reply_start) with jobId/threadId handling, a concrete JSON schema for REVIEW_STATE.json, a copy-paste markdown template for documenting rounds, a literal human-checkpoint message template, and explicit input-parsing rules ("go"/"continue"/"skip 1,3"/"stop"). This matches the 5 anchor — copy-paste-ready commands and templates covering the common cases.

Workflow Clarity

The sequence is explicit — Initialization (with resume/stale-state validation branches) → Loop Phases A-E → Termination — with an exact STOP CONDITION (score >= 6 AND verdict in {"ready","almost"}), a state file written every Phase E for recovery, and a true feedback loop (review → implement fixes → re-review). This matches the 5 anchor: clear sequencing with explicit validation checkpoints and error-recovery handling.

Progressive Disclosure

The body is well-sectioned with clear headers, but it is a single ~300-line file with no bundle files (references/, scripts/, assets/ absent), and large blocks that belong in a reference are inlined — the duplicated SCOPE LIMITS prompt text and both full prompt templates. The Output Protocols links to ../../shared-references/*.md are one level deep but those files are not present in this bundle, so navigation cannot be verified. It is not 2 because structure and signaling are decent, and not 4 because significant content that should be split out is inline.

Specificity

The description lists several concrete actions — "Repeatedly reviews using Gemini via gemini-review MCP, implements fixes, and re-reviews until positive assessment or max rounds reached" — which is concrete and actionable. It is not 5 because coverage is partial (state persistence, checkpoints, and notifications are unmentioned), and not 3 because it goes well beyond 1-2 generic actions to name the mechanism, the reviewer, and the termination condition.

Completeness

It explicitly answers both questions: what ("Autonomous multi-round research review loop... implements fixes, and re-reviews until positive assessment or max rounds reached") and when ("Use when user says 'auto review loop', 'review until it passes', or wants autonomous iterative improvement") with concrete trigger phrases. This matches the 5 anchor exactly; neither what nor when is vague or implied.

Trigger Term Quality

Triggers are natural phrases a user would actually say, quoted verbatim: "auto review loop", "review until it passes", "wants autonomous iterative improvement". It is not 5 because common synonyms and variations (e.g. "iterate until it passes", "keep reviewing", "make the review pass") are missing, and not 3 because multiple genuine user-facing phrases are present rather than generic keywords only.

Distinctiveness Conflict Risk

Naming "Gemini via gemini-review MCP" and "auto review loop" carves a clear niche that is unlikely to fire for unrelated skills, though the broad words "review" and "iterative improvement" leave minor overlap risk with general code-review/PR-review skills. It is not 5 because of that overlap, and not 3 because the MCP bridge and quoted trigger phrases make it substantially distinguishable.