Content
63%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, domain-rich pipeline: clear phase sequencing, named tools with parameters, a decision matrix, quality checklist, and thoughtful edge-case fallbacks. It loses points on conciseness (repeated resolution guidance, a basic EM-modality primer, a tangential IDR section) and on progressive disclosure, where a ~290-line monolith inlines tool API details and an off-topic related-resource guide that belong in reference files.
Suggestions
State the resolution bands once (Phase 1's interpretation table is the best home) and have the guiding principles and decision matrix reference them instead of restating them.
Move the per-tool input/output documentation for Phases 1-5 and the entire IDR section into references/ files (e.g., references/tools.md and references/idr.md), keeping SKILL.md as a phase-and-workflow overview with one-level-deep pointers.
Drop or compress the 'TEM resolves individual protein complexes (~2nm)... SEM shows surface topology' paragraph — it is textbook EM knowledge Claude already has — and add one runnable tu.run_tool example per phase so tool invocation syntax is explicit.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient domain-specific content (tool names, FSC 0.143 vs 0.5 cutoff, cross-correlation > 0.7 thresholds), but with real tightening opportunities: resolution bands are stated three times ('sub-4A enables atomic modeling, 4-8A enables domain fitting, >8A is shape-level' in the principles, the Phase 1 'Resolution interpretation' table, and the Phase 6 decision matrix), the 'EM resolution determines what you can see. TEM resolves... SEM shows surface topology' paragraph is basic knowledge Claude already has, and the ~25-line IDR light-microscopy section is tangential to an EM skill. This matches 'Mostly efficient but includes some unnecessary explanation or could be tightened' rather than a 4, where only minor trimming would be needed. | 3 / 5 |
Actionability | Concrete, callable guidance: every phase names its tools with inputs and outputs ('EMDB_search_structures -- Input: query, optional resolution_min, resolution_max, method, limit'), and the IDR section even gives a runnable invocation ('tu.run_tool("IDR_search_studies", {"key": "Gene Symbol", "value": "TP53"})'). Not a 5 because the EM/PDB/CryoET tool listings stop at parameter tables — none include an actual call example in the tu.run_tool form that the IDR section demonstrates, leaving the exact invocation syntax implicit for ~15 tools. | 4 / 5 |
Workflow Clarity | A clearly sequenced 7-phase pipeline with per-phase numbered workflows, a decision matrix ('Which map should I use?'), a quality assessment checklist, and an Edge Cases & Fallbacks section covering failure paths ('No EMDB entries' → search PDB with method filter). Not a 5: while the Phase 6 checklist and fallback branches approximate validation, there are no explicit verify-then-retry loops (e.g., confirming a returned EMDB entry actually matches the queried target before building on it), and not a 3 since checkpoints like the FSC-cutoff and fitting-statistics checks are explicit rather than absent. | 4 / 5 |
Progressive Disclosure | The skill is a single ~290-line file with no bundle files at all; the per-tool API documentation (inputs/outputs for ~15 tools across Phases 1-5) and the entire IDR section are inline material that would naturally live in references/ files. Section headers and tables give it real structure — above anchor 2's '300 lines of API reference with no section headers' — but it matches anchor 3: 'Some structure... content that should be separate is inline'. Not a 4/5: the guideline's exception for well-organized single files applies to skills under 50 lines; this one has clear separation candidates it doesn't use. | 3 / 5 |
Total | 14 / 20 Passed |