Content
75%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 skill body with a clear four-phase workflow and concrete MCP tooling. It is held back from the top band by minor verbosity, unspecified review-failure recovery, and the lack of any progressive-disclosure file split for its report template.
Suggestions
Trim rhetorical padding such as "Be BRUTALLY honest — false novelty claims waste months of research time" to a terse directive, since the instruction already conveys the requirement.
Add an explicit error-recovery branch for Phase C (e.g., what to do if review_status returns an error or never reaches done=true within a bounded retry count) to close the workflow-validation gap.
Move the full Novelty Check Report markdown template into a references/ file and link to it from Phase D, keeping the inline overview lean and giving the skill a one-level-deep progressive-disclosure structure.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is efficient and assumes Claude's competence (no explainer on what arXiv or Scholar is), but a few rhetorical lines like "Be BRUTALLY honest — false novelty claims waste months of research time" could be trimmed, so it sits above the mostly-efficient anchor and below the lean-every-token-earns-its-place anchor. | 4 / 5 |
Actionability | It gives concrete executable guidance including the actual MCP tool call (mcp__gemini-review__review_start), a bounded polling pattern (waitSeconds until done=true), and a copy-paste report template, but Phase B leaves exact query strings to Claude, leaving minor gaps that keep it below fully-executable 5. | 4 / 5 |
Workflow Clarity | Phases A through D are clearly sequenced and Phase C includes an explicit polling checkpoint (poll review_status until done=true), matching the clear-sequence-with-most-checkpoints anchor; it falls short of 5 because error-recovery behavior if the review job fails is not specified. | 4 / 5 |
Progressive Disclosure | The single-file skill is well-organized into Constants, Instructions (Phase A-D), and Important Rules sections with no nested references, fitting the good-structure anchor; at ~80 lines it exceeds the under-50-line simple-skill exception that would allow a 5, and the inline report template could arguably live in a separate file. | 4 / 5 |
Total | 16 / 20 Passed |