Content
60%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.
The skill has a clear sequenced workflow, a concrete mandatory output template, and well-signaled one-level-deep references that all exist on disk. Its main weakness is heavy redundancy — the same failure modes and rules are restated in four or more sections and duplicated from the reference files — which roughly doubles the token cost without adding guidance.
Suggestions
Collapse the redundant sections — 'Task', 'Core Function', 'Hard Rules', 'What This Skill Should Not Do', and 'Quality Standard' restate the same mismatch/overextension/second-hand/quote-drift list; keep one canonical statement and point to the reference files for the rest.
Remove the inline 'Hard Rules' section (or reduce it to a one-line pointer) since references/hard-rules.md already carries those rules, and likewise the severity-level list duplicated in severity-classification-rules.md.
Add one brief worked example — a claim-reference pair showing an overextension and how it would be reported under the A-H output structure — to lift actionability without much token cost.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body repeatedly restates the same failure-mode list — mismatch, overextension, second-hand citation, quote drift — across the intro bullets, 'Task', 'Important Distinctions', 'Core Function', Steps 3-5, 'Hard Rules', 'What This Skill Should Not Do', and 'Quality Standard'; 'Input Validation' duplicates 'Step 1', and 'Hard Rules' duplicates references/hard-rules.md nearly verbatim. Several sections are pure padding that could be cut by half without losing guidance, matching the 'noticeably verbose; several padded sections' anchor rather than the mostly-efficient one. | 2 / 5 |
Actionability | For an instruction-only skill the guidance is concrete and specific: a mandatory eight-part output template (A-H), explicit matching criteria ('same population... same intervention/exposure/context... same evidence level... same direction of claim'), a named severity taxonomy, and concrete correction actions ('replacement citation, softer wording, narrower claim language, citation relocation'). It misses a 5 because there are no worked examples — no sample claim-reference pair showing a mismatch or a sample structured output, which the rubric's good examples treat as covering common cases. | 4 / 5 |
Workflow Clarity | Nine explicitly sequenced steps with a validation gate up front (Step 1 plus output section 'A. Input Match Check') and a mandatory output structure acting as a checklist. It does not reach 5 because there is no explicit feedback loop for what to do after requesting missing material or re-checking after user response — the recovery path is only implied by section H. | 4 / 5 |
Progressive Disclosure | Seven reference files all exist in references/ and are each one level deep with no nested references, and the 'Reference Module Integration' section clearly signals each file's purpose and when to apply it ('Use before any long-form integrity review', 'Apply throughout the entire response'). It stays below 5 because the body inlines content duplicated in those files (the ten 'Hard Rules' vs references/hard-rules.md, the severity levels vs severity-classification-rules.md), so the split is not fully clean. | 4 / 5 |
Total | 14 / 20 Passed |