Content
71%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, highly actionable audit workflow with excellent progressive disclosure into a genuine one-level-deep bundle. Its main weakness is redundancy — the same do's and don'ts are repeated across four sections — which inflates token cost without adding guidance, and the workflow lacks an explicit feedback loop after the final self-check.
Suggestions
Consolidate the overlapping directive sections ('Core Function' should-not list, 'Hard Rules', 'What This Skill Should Not Do', 'Quality Standard') into a single rules section to remove triple- and quadruple-stated anti-patterns.
Add a feedback loop after Step 8: if the self-critical check exposes a missed fragility or unsupported judgment, revisit the relevant earlier step before finalizing.
Include one brief worked example (claim → evidence chain → judgment) in the body or a reference to anchor the interpretive steps that currently rely on judgment alone.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is composed of dense, directive instructions with no explanation of concepts Claude already knows, but the same anti-patterns are stated three to four times across 'Core Function', 'Hard Rules', 'What This Skill Should Not Do', and 'Quality Standard' (e.g. 'do not treat high performance metrics as reliability by default' / 'Do not equate high AUROC / C-index / accuracy with robustness'), and the skill's purpose is restated in four sections. This is 'mostly efficient but could be tightened' rather than 2, since the padding is structural duplication of directives, not padded explanation; it is not 4 because the redundant sections are a real, multi-section trimming opportunity. | 3 / 5 |
Actionability | The body gives concrete, executable audit guidance: 8 ordered steps with specific per-step checklists ('events-per-variable', 'calibration vs discrimination', 'multiple testing handling'), a mandatory A–I output structure, a verified table format in the reference module, a ready-to-use out-of-scope redirect template with a [restatement] placeholder, and concrete example inputs and sample triggers. It falls just short of 5 because several directives remain interpretive ('what the paper is asking the reader to believe', 'most likely hidden fragility') without worked examples, and the judgment-level definitions are delegated to a reference file. | 4 / 5 |
Workflow Clarity | The 'Execution — 8 Steps (always run in order)' provides an explicit sequence with real checkpoints: gating ('Do not evaluate reliability before the claim-to-evidence chain is explicit'), per-step reference dispatch ('Apply references/design-and-bias-rules.md'), and a step-8 self-critical final check. It is not 5 because there is no feedback loop — nothing instructs revisiting an earlier step if the final check exposes a missed fragility — and the step-8 check is a one-pass review rather than an explicit validate-and-recover cycle. | 4 / 5 |
Progressive Disclosure | The body acts as a clear orchestration overview: all seven referenced files exist in references/, each is listed in 'Reference Module Integration' with a one-line purpose ('Use for the core audit dimensions...', 'Use when checking design fit...'), each is dispatched at a defined execution step, and the references are exactly one level deep (none of them point onward). Content is appropriately split — frameworks, rule detail, and output guidance live in the bundle while the body holds sequencing and orchestration. | 5 / 5 |
Total | 16 / 20 Passed |