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.
The body is a dense, highly actionable reference with real verified tool behavior that Claude could not infer — its greatest strength — but it is a monolithic 317-line file with no progressive disclosure, systematically repeated gotchas, and verbose narrative examples. Actionability is top-tier; conciseness and file structure are the clear improvement areas.
Suggestions
Deduplicate the gotchas: state each once in its workflow (or once in the Gotchas table) instead of the current threefold repetition of PDBe_get_compound_structures, MobiDB unreachability, and InterPro flakiness across workflow, table, and Limitations sections.
Move the Tool Inventory section (~55 lines of tool signatures) into a references/tool-inventory.md file and keep SKILL.md to workflows, decision rules, and interpretation tables, cutting the main file roughly in half.
Tighten the 'Real example' narratives to the call plus the key output fields (e.g. reduce the MDDB example to the tool call, project accession, and one-line result), since the extended parameter dumps repeat information already in the tool inventory.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body mostly earns its tokens with verified-live tool behaviors Claude cannot know (parameter gotchas, the MobiDB endpoint hang, the PDBe_get_compound_structures fallback), but several gotchas are stated three times each (compound_structures in Workflow 2, the Gotchas table, and Limitations; likewise MobiDB unreachability and InterPro flakiness), and the 'Real example' narratives run long (e.g. the full MDDB parameter dump for MD-A001UA). This is 'mostly efficient but includes some unnecessary explanation or could be tightened' — beyond the score-4 anchor's 'minor instances', but well short of the severe padding of score 2. | 3 / 5 |
Actionability | Guidance is fully executable: exact tool names with parameter signatures ("alphafold_get_prediction (qualifier=UniProt)"), copy-paste-ready verified calls with real arguments and outputs ("InterPro_get_member_signature(member_database='pfam', accession='PF00069')"), and a Mistake/Correct gotcha table. Specific examples cover the common cases (kinases, p53, rhodopsin, lysozyme, ubiquitin, ATP/HEM), matching the score-5 anchor; it does not fall to score 4 because no execution gaps remain for an agent with these tools. | 5 / 5 |
Workflow Clarity | The nine workflows are clearly phased with tool-per-phase sequences, and several include explicit validation checkpoints ("PDBeValidation quality → binding site well-resolved?", BMRB_get_validation gating shift trust) plus a retry feedback loop for the flaky InterPro endpoints. However, checkpoints are uneven — Workflows 1-3 phases are terse one-liners ('Phase 8: Evidence integration') with implicit validation — so it fits 'Clear sequence with most checkpoints present; minor validation gaps' rather than the score-5 anchor requiring explicit validation and feedback loops throughout. | 4 / 5 |
Progressive Disclosure | The section structure is good (LOOK UP DON'T GUESS, tool inventory, workflows, gotchas, evidence grading, limitations), but the skill is a 317-line monolith with zero reference files: the ~55-line tool inventory and the long verified-example narratives are exactly the content the rubric says belongs in a separate references/ file. This matches 'Some structure... content that should be separate is inline' rather than score 4 ('most content is appropriately placed'), since nothing is split out at all; it stays above score 2 because internal headers make navigation genuinely easy. | 3 / 5 |
Total | 15 / 20 Passed |