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.
This is a highly actionable, well-sequenced skill body with excellent tool-level specificity, executable code, and built-in validation/feedback loops. Its weaknesses are length and self-containment: the trap summaries are duplicated in later sections and ~340 lines of per-database and computational detail are inlined in SKILL.md rather than split into one-level-deep reference files.
Suggestions
Split the self-contained computational procedures (ORF counting, restriction digest, segregation ratios) and the per-database deep dives (Human Protein Atlas, Allen Brain, FDA-label sections, gnomAD release table) into reference files under references/, leaving SKILL.md as a routing-table overview with one-level-deep pointers — this would materially cut token cost while preserving actionability.
De-duplicate the six-trap summary against the detail sections: keep one-line reminders in the trap list and move the worked examples (rs2475055 p-values, #F2483B, APOC2 pLI) to the single detailed section only, rather than restating each trap twice.
Trim rhetorical asides ('a guess wearing a citation', 'resist it', 'do not let this happen') to plain imperatives — the instruction lands the same with fewer tokens.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is information-dense and avoids explaining concepts Claude already knows, but it could be tightened: the six traps are summarized up top and then each is re-explained at length below ('HPA subcellular locations pool every cell line' vs. the full 'Human Protein Atlas — report both location fields' section; the Allen Brain and gnomAD release items likewise), and some rhetorical padding ('a guess wearing a citation') adds tokens without adding instruction. Anchor 3 ('mostly efficient but... could be tightened') fits better than 4, where only minor trimming would be needed. | 3 / 5 |
Actionability | Fully executable throughout: exact tool names with parameters (`MSigDB_get_gene_set_members`, `DisGeNET_get_disease_genes(disease=CUI)`, `SCREEN_search_cCREs_by_region`), exact set-name conventions (`MIR186_3P`, `MP_INCREASED_MELANOMA_INCIDENCE`), and three complete, runnable Python snippets (ORF counting, restriction digest, gamete ratios) with worked examples covering the common cases. | 5 / 5 |
Workflow Clarity | Multi-step processes are explicitly sequenced with validation checkpoints and feedback loops: the MCQ procedure (Parse → Resolve → Query → Check → Answer), the 6-step differential gene–disease recipe, and the drug-vignette steps each embed error recovery ('If a tool errors on a name, re-derive the name... before concluding insufficient'; 'If RestrictionBatch raises on an enzyme name... re-run — do not fall back to guessing') and validation ('Always check the returned length against what the question asked for'; 're-read each option against the computed result before emitting [ANSWER]'). | 5 / 5 |
Progressive Disclosure | The body is well-sectioned with clear headers and internal cross-links, but it is a ~340-line monolith with no bundle files at all: content that clearly belongs in separate reference files — the three full computational procedures with code, the per-database deep dives (HPA, Allen Brain, FDA-label table, gnomAD release table) — is inlined. This sits between anchor 2 (content that belongs in separate files is inlined) and anchor 3 (some structure, could be better organized), but the section structure is genuinely good, so 3 rather than 2. | 3 / 5 |
Total | 16 / 20 Passed |