Content
67%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A well-structured, actionable instruction skill with a clear phase workflow, concrete MCP integration details, and strong calibration guidance. The main cost is redundancy — the verdict-limits philosophy is stated twice and padded — and a missing error path for a failed review job.
Suggestions
Deduplicate the "two failures waste months equally" thesis and the proximity/ABANDON reasoning, which appear in both the NOVELTY VERDICT LIMITS block and Important Rules — state each idea once and cross-reference it.
Add 2-3 example search query formulations in Phase B (e.g., for a method claim: "<technique> + <task> arXiv 2025") so the "3 different query formulations" rule is copy-paste executable rather than abstract.
Add a fallback path in Phase C for when the review job errors or times out (e.g., bounded retry count, then report Phase B findings directly with a note that independent review was unavailable).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | No padding of concepts Claude already knows, but the body repeats itself: "Two failures waste months equally" and its false-positive/false-negative thesis appear in both the NOVELTY VERDICT LIMITS block and Important Rules, and the verdict limits are philosophically repetitive ("CAUTION is not a safe middle", "choose the more permissive one"). It could be tightened without losing content; not 4 because the duplication is more than minor trimming. | 3 / 5 |
Actionability | Concrete, executable guidance throughout: named MCP tools (mcp__claude-review__review_start / review_status), the polling protocol (save jobId, bounded waitSeconds until done=true, retain threadId), 3 query formulations per claim with year filters, WebFetch of abstracts, and a full output template. Not 5: the MCP call shows a placeholder prompt rather than a copy-paste-ready example, and no sample search queries are given. | 4 / 5 |
Workflow Clarity | A clearly sequenced Phase A→D workflow with a verification checkpoint (independent reviewer with poll-until-done) and explicit decision rules for the verdict. Not 5: there is no error-recovery path if the review job fails or times out, and no fallback if the reviewer is unreachable. | 4 / 5 |
Progressive Disclosure | A single well-sectioned file with clear headers (Constants, Instructions, phases, Important Rules, Review Tracing) and one clearly-signaled one-level reference (../shared-references/review-tracing.md) with its output directory specified. The verdict-limits block and report template are rightly inline since they must be copied verbatim; no bundle files exist to split into. Not 5: the external reference is outside the bundle and the skill is monolithic beyond the ~50-line simple-skill case. | 4 / 5 |
Total | 15 / 20 Passed |