Content
56%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 tool-facing core is excellent — verified live examples, parameter tables, and real gotchas (Complex Portal's null echoed complex_id, IntAct's metadata-nested interaction_ids). However, the skill is dragged down by ~80 lines of textbook binding-thermodynamics/coiled-coil/detergent material Claude already knows, a stray memorized Q&A passage, and references to bundle files that don't exist.
Suggestions
Remove or move to a reference file the domain-textbook sections Claude already knows (Statistical Factors in Multimeric Binding, binding-curve-vs-stoichiometry table, Coiled-Coil Prediction, Detergent Effects), keeping only the tool-driven 'LOOK UP DON'T GUESS' guidance.
Actually provide python_implementation.py and KNOWN_ISSUES.md in the bundle (or drop the references), since the body directs the reader to them for runnable examples and verbose-output filtering.
Delete the 'Amyloid clearance... The answer is collagen' and 'give JUST the protein name' passages — they are answers to specific past questions, not reusable skill guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Several sections re-teach textbook biochemistry Claude already knows: 'Statistical Factors in Multimeric Binding' (kon/koff factors, Kd_macro vs Kd_micro), the ITC/SPR/Hill/Scatchard comparison table, 'Coiled-Coil Oligomeric State Prediction' (heptad repeats, a/d positions), and 'Detergent Effects on Membrane Proteins' (DDM/LMNG vs SDS). The 'Amyloid clearance... The answer is collagen' passage reads as a memorized Q&A answer rather than skill guidance. These pad an otherwise lean tool guide, making it noticeably verbose overall — though the tool-specific sections themselves are tight and high-value. | 2 / 5 |
Actionability | Concrete, executable guidance dominates the tool sections: real verified calls like 'SIGNOR_connect_proteins(proteins=["P00533", "P28482"], level=2)', 'ComplexPortal_search_complexes(query="TP53", species="Homo sapiens")', and 'EBIProteins_get_interactions(accession="P04637")' with parameter tables and a parsed-response snippet. Minor gaps: the core 4-phase tools (STRING_map_identifiers, STRING_get_network) are named without example calls, deferring to python_implementation.py — which does not exist in this bundle. | 4 / 5 |
Workflow Clarity | The 4-phase workflow (identifier mapping → network retrieval → enrichment → optional structural) is clearly sequenced, with a Complex Portal two-phase lookup, two worked SIGNOR workflows, and a troubleshooting section covering failure modes (no interactions found → lower confidence_score; BioGRID failing → set API key). Operations are read-only, so no destructive-validation cap applies; checkpoints are present but mostly implicit (verify names, check escore, confirm complex_id) rather than explicit validate-fix-retry loops. | 4 / 5 |
Progressive Disclosure | Section structure is decent (headers, parameter tables, per-database notes) and references are clearly signaled inline, but the bundle contains no references/, scripts/, or assets/ — so 'See python_implementation.py for runnable examples' and '(see KNOWN_ISSUES.md)' are dangling paths. Meanwhile ~80 lines of binding-thermodynamics and structural-biology domain reasoning that belong in a separate reference file are inlined in SKILL.md. | 3 / 5 |
Total | 13 / 20 Passed |