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 detailed, actionable, well-sequenced workflow with real validation and feedback loops, undermined by duplicated content and an internal contradiction about the Round 2 review tool. Structure is good but bulkier than necessary for a SKILL.md overview.
Suggestions
Resolve the Round 2 contradiction: Step 5 says to start a fresh review_start while Key Rules says to use review_reply_start to maintain context — pick one and align both sections.
De-duplicate the ~25-line claim-calibration review prompt: factor it into a single reference (e.g. references/review-prompt.md) and have Step 2 and Step 5 point to it, or define it once and reference by variable.
Remove or reconcile the 'Typical Score Progression' Round 3 row, which conflicts with MAX_ROUNDS = 2 stated in Constants.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient domain-specific guidance, but the long claim-calibration review prompt is duplicated nearly verbatim in Step 2 and Step 5, polling instructions are repeated, and the 'Typical Score Progression' table invents a Round 3 that contradicts MAX_ROUNDS=2. | 3 / 5 |
Actionability | Concrete copy-paste bash (cp, latexmk, pdfinfo, grep) and full MCP prompt templates make the guidance mostly executable, but placeholders like [VENUE] and a contradiction over which MCP tool to use in Round 2 leave minor gaps. | 4 / 5 |
Workflow Clarity | Steps 0–9 are clearly sequenced with explicit validation ('Verify: 0 undefined references, 0 undefined citations') and a format-check feedback loop, but Step 5 ('start a fresh review_start — do not reuse the Round 1 thread') contradicts Key Rules ('Use review_reply_start ... for Round 2 to maintain conversation context'), creating real ambiguity. | 4 / 5 |
Progressive Disclosure | Section headers are clear and shared protocols are signaled via one-level-deep links (../../shared-references/output-versioning.md etc.), but the large inline review-prompt and fix-table blocks—some duplicated—could live in reference files; no bundle files are present. | 4 / 5 |
Total | 15 / 20 Passed |