Content
73%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-documentation core is excellent: concrete, verified against live data, with clear workflows, gotchas, and an explicit resource-selection guide. The main weaknesses are Sections 1-4 restating textbook neuroscience Claude already knows, and the absence of any code example despite mandating Python for computations.
Suggestions
Cut or move Sections 1-4's textbook material (cranial nerve exam, UMN/LMN table, action-potential phases, basal-ganglia pathways) to a reference file, keeping only the verification prompts and the pitfalls from Section 12 — Claude already knows these facts.
Add one compact executable Python example supporting the 'ALWAYS use Python' mandate, e.g. fixed-point iteration for the steady-state population-rate equation in Section 1.
Add example queries for the Section 5 tools that lack them (DANDI, ClinVar, GWAS, Orphanet), mirroring the verified-example pattern used in Sections 7-10.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The tool sections (5, 7-11) are dense and every token earns its place — verified live examples, gotchas, and call-chain workflows are knowledge Claude cannot reconstruct. However, Sections 1-4 (~130 of 357 lines) restate textbook neuroscience Claude already knows (cranial nerve exam, UMN vs LMN table, action-potential phases, basal-ganglia pathways, Nernst equation). Mostly efficient with several unnecessary reference sections; sits between anchors 2 and 3 because the padding is compact bullet format rather than verbose prose. | 3 / 5 |
Actionability | Tool guidance is highly concrete: named tools with key parameters, ready-to-use query patterns ('[region A] projection [region B]', '[region] lesion', '[function] brain region fMRI'), and verified live examples with real IDs (ds000117, model 3263, collection 457, neuron 147055). Falls short of 5 because the 'ALWAYS use Python for multi-step calculations' mandate includes no executable code example, and several Section 5 tools (DANDI, ClinVar, GWAS, Orphanet) have no example query. | 4 / 5 |
Workflow Clarity | Multi-step processes are clearly sequenced with explicit validation checkpoints: numbered reasoning patterns for 'where in the brain' and clinical questions (steps 1-5 each), quantitative problem-solving with unit and sanity-check steps, four resource-specific tool call chains, and error-recovery gotchas ('Don't treat an empty edges array as ... no datasets exist', the NeuroVault non-working-search caveat, and the base64-image warning). Matches anchor 5's explicit validation and feedback-loop criteria. | 5 / 5 |
Progressive Disclosure | No bundle files exist, so everything is inline in a single 357-line SKILL.md. Structure is good: numbered, well-labeled sections; cross-references between them (Sec. 11 points back to Secs. 5, 7-10); and a resource-selection comparison table aids navigation. Not 5 because reference-style material (the full cranial-nerve/clinical atlas in Sec. 3 and the per-tool tables) is inlined where one-level-deep reference files would better serve token economy; not 3 because navigation is genuinely easy and clearly signaled. | 4 / 5 |
Total | 16 / 20 Passed |