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 routing skill with a concrete dispatch table, specific tool/field guidance, and clear sequenced steps with verification gates. Its main defect is that the body depends on seven reference files that are absent from the bundle, breaking the progressive-disclosure path and leaving the actual diagnostic procedures unreachable.
Suggestions
Ship the references/ directory: the 7 referenced files (diagnostic-snapshot.md, bias-and-skew.md, empty-experiment.md, interpretation.md, numbers-vs-sql.md, mid-run-changes.md, qualitative-feedback.md) are required for the skill to function — without them every diagnostic group routes to a missing file.
Tighten Step 3's 'don't collapse mechanisms' paragraph and Step 4's reversal-offer paragraph; the core guidance (surface co-occurring mechanisms independently; don't preemptively offer reversal) can be stated in roughly half the tokens.
Add a brief inline fallback for the most common diagnostics (e.g. the 0-exposure / SRM quick checks) so the skill remains useful even before a reference file is opened, reducing single-point dependence on the missing bundle files.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient: a concrete dispatch table and a specific experiment-get field list that assume Claude's competence, but the Step 3 'don't collapse mechanisms' rationale and the Step 4 reversal-offer paragraph are lengthier than needed and could be trimmed — between anchors 3 and 5, leaning up. | 4 / 5 |
Actionability | Gives concrete executable guidance for the routing layer ('Call experiment-get and pull these fields', a complete symptom→group table, explicit state-scoping rules), but the actual diagnostic procedures are delegated to reference files — mostly executable with the gap that deep steps live elsewhere. | 4 / 5 |
Workflow Clarity | A clearly numbered sequence (Step 1 → 1.5 → 2 → 3 → 4) with verification gates ('verify before asking', 'only list mechanisms with a path to verification', state-calibration), but the deeper validation/feedback loops reside in the unshipped reference files, leaving minor gaps in the body itself. | 4 / 5 |
Progressive Disclosure | The design is excellent — each diagnostic group A–F summarizes then signals a one-level-deep '→ See references/X.md' — but the bundle ships no references/ directory, so the 7 referenced detail files (diagnostic-snapshot, bias-and-skew, empty-experiment, interpretation, numbers-vs-sql, mid-run-changes, qualitative-feedback) do not exist; the overview points to materials that aren't shipped, so disclosure doesn't actually resolve. | 3 / 5 |
Total | 15 / 20 Passed |