Content
77%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 instruction skill with excellent workflow sequencing, validation checkpoints, and clean one-level progressive disclosure into eight real reference files. The main weakness is redundancy: the Hard Rules section substantially duplicates the step sections, inflating the token budget without adding new guidance.
Suggestions
Consolidate the 23 Hard Rules: keep only rules not already stated in Steps 1–7 (e.g., DESeq2 vs limma choice, redirect scope) and convert the rest into one-line cross-references to the step sections, cutting the body by roughly a quarter.
Add one compact worked example (a filled-in Standard-plan outline for a single trigger like the diabetic-nephropathy example) so the A–J output contract is unambiguous on first use.
Merge the duplicated dependency-consistency text (Step 5 checklist vs. Hard Rules 16/18/19 and the C.5 section description) into a single canonical checklist with pointers.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body mostly avoids explaining concepts Claude already knows and stays instructional, but it is noticeably repetitive and could be tightened: the 23 Hard Rules restate step-section content almost verbatim (rules 15/17/20 restate Section G and C.5 mandates; rules 11–13 restate Step 4.5's citation rules; rule 22 restates Section D's disclaimer instruction; rules 16/18/19 restate Step 5's dependency checks). This matches 'mostly efficient but includes some unnecessary explanation or could be tightened'. Not a 2: there is no domain-knowledge padding (no library tutorials, no PDF-style concept explanations) — the excess is duplicated rules, not filler. Not a 4: the duplication is substantial and systematic, not a minor instance. | 3 / 5 |
Actionability | For an instruction-only skill the guidance is highly concrete: an exact input format ('[disease / condition] + [process-related gene family / pathway / biological theme] + [validation direction]'), worked example inputs, a copy-paste redirect message template, the exact Dataset Disclaimer line to emit, explicit dependency formulas, and a fully specified A–J output section structure. It is not a 5 because there is no worked example of a complete (even skeletal) output for one concrete case, and a few directives remain abstract ('Explain the logic of each decision', 'Explain why this is the best match') with the actionable 8-field template deferred to a reference file. It is clearly above 3: the guidance goes far beyond pseudocode-level hints. | 4 / 5 |
Workflow Clarity | The 'Execution — 7 Steps (always run in order)' gives a clear mandatory sequence with an explicit validation checkpoint (Step 5 'Dependency Consistency Check (mandatory before output)') and a feedback loop ('If any dependency inconsistency is found, revise the plan before outputting'). Additional checkpoints and recovery paths are present: STOP-and-redirect on out-of-scope, state-assumptions fallback for insufficient input, and the literature fallback ('If browsing/search is unavailable... output a search strategy + target evidence map instead of fake references'). This matches the top anchor: clear sequence, explicit validation, error-recovery loops. | 5 / 5 |
Progressive Disclosure | SKILL.md functions as an overview with eight clearly signaled, one-level-deep references ('→ Detailed pattern logic: [references/study-patterns.md]', etc.), all of which exist in references/ and contain real, distinct content (verified against the file listing and workload-configurations.md). Detail is appropriately split out (config tables, method library, step template, figure plan, evidence hierarchy) with a concise decision skeleton left inline. Not below 4: every reference is explicit, shallow, and purpose-labeled, and no reference chain nests further. | 5 / 5 |
Total | 17 / 20 Passed |