Conciseness

The body is mostly operational, but there is noticeable padding that could be tightened: the cross-family assurance appears twice (the opening blockquote and "this route is a different model family from the Codex executor and records review_independence: cross-family" restated in the overlay assurance note), the polling instruction "immediately save the returned jobId and poll mcp__claude-review__review_status with a bounded waitSeconds until done=true" is repeated verbatim in Steps 2 and 3, and the ~25-line "SCOPE LIMITS" policy block plus the Prompt Templates section restate material. Anchor 3 (mostly efficient but could be tightened) fits; not 2 because there is no explanation of concepts Claude already knows and most tokens are functional.

Actionability

Concrete, executable guidance throughout: the install command `codex mcp add claude-review -- python3 ~/.codex/mcp-servers/claude-review/server.py`, exact tool names (`mcp__claude-review__review_start`, `review_reply_start`, `review_status`), fully written prompt bodies, and `tools: "Read,Grep,Glob"`. Minor gaps keep it below anchor 5: `threadId: [saved reviewer id from Step 2]` and `[Full research context + specific questions]` are placeholders, and there is no concrete example of the review_status polling call or a suggested waitSeconds value. Well above anchor 3 — this is real copy-paste-ready material, not pseudocode.

Workflow Clarity

Six clearly sequenced steps (gather context → initial review → iterative dialogue → convergence → document → trace) with explicit validation checkpoints ("poll mcp__claude-review__review_status with a bounded waitSeconds until done=true") and explicit stop criteria in Step 4. Anchor 4 fits: minor validation gaps — no guidance on what to do if a job errors, stalls, or never reaches done=true. Not 5 because there is no error-recovery feedback loop; not 3 because polling validation and convergence conditions are explicit, not implicit.

Progressive Disclosure

The body is well-sectioned (Constants, Prerequisites, Workflow, Key Rules, Prompt Templates) and the two references ([output-composition.md](../shared-references/output-composition.md) and ../shared-references/review-tracing.md) are clearly signaled one level deep. Anchor 4 fits: minor organization gaps — the referenced shared files live outside this skill's directory with no bundle files shipped alongside, and the long SCOPE LIMITS block and prompt templates are inlined where a reference file would reduce SKILL.md weight. Not 3 because the references that do exist are clearly signaled, and the overall structure is good; not 5 because content that could be split out remains inline and no in-bundle reference organization exists.

Specificity

"Get a deep critical review of research" names the domain and one concrete action applied to targets ("research ideas, papers, or experimental results"), but only that single review action is stated — not the multi-round dialogue, experiment design, or mock-review capabilities the skill actually has. This matches anchor 3 (domain and 1-2 concrete actions, not comprehensive), not anchor 4 which requires several listed specific actions.

Completeness

Explicitly answers both: what ("Get a deep critical review of research from Claude via claude-review MCP") and when ("Use when user says \"review my research\"... or wants critical feedback on research ideas, papers, or experimental results") with concrete trigger phrases. Matches anchor 5 exactly; not 4 because the when-clause is fully explicit, not merely present.

Trigger Term Quality

Direct user-voice phrases "review my research", "help me review", "get external review", plus synonyms "critical feedback" and target terms "research ideas, papers, or experimental results" give comprehensive natural-term coverage. Not 4: it captures the actual sentences a user would say, including variations, rather than leaving common synonyms missing.

Distinctiveness Conflict Risk

The research-review niche and claude-review MCP routing are distinct, but "help me review" and "get external review" could also match generic code-review or PR-review skills. Anchor 4 (mostly distinct, minor overlap risk with closely related skills) fits; not 5 because of that overlap with general review skills, and not 3 since the research framing is consistent throughout.