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.
The tool-facing half of this skill is excellent — concrete tool tables, chained workflows, validation checkpoints, and verified live examples with real IDs. The main weakness is the first half: roughly 150 lines of textbook neuroscience (plasticity rules, cranial nerves, Nernst equation) that Claude already knows, inlined into SKILL.md rather than split into reference files, which inflates token cost without adding skill-specific value.
Suggestions
Move the textbook reference material (sections 1-4: computational models, neuroanatomy divisions, cranial nerve exam, neurophysiology equations) into a references/ file such as references/neuroscience-facts.md, keeping only the tool-specific lookup guidance and reasoning patterns in SKILL.md.
Trim the triple-stated 'LOOK UP, DON'T GUESS' principle to a single statement, and cut general-knowledge lines like the Hebbian/BCM definitions and AP phase descriptions that Claude already knows.
Add a small executable Python example to the quantitative problem-solving strategy (e.g. a fixed-point iteration for r = f(W*r + I_ext)), so 'ALWAYS use Python' is backed by copy-paste-ready code like the tool sections are.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Sections 1-4 (~150 lines) extensively explain concepts Claude already knows: STDP windows, Hebbian/BCM theory, the full cranial-nerve exam table, UMN vs LMN table, Nernst/Goldman equations, basal ganglia direct/indirect pathways, and action-potential phases. The 'LOOK UP, DON'T GUESS' principle is stated three times (key principles, section header, and section body). This matches anchor 2's 'noticeably verbose; several unnecessary explanations or padded sections' rather than anchor 3, because the majority of the first half is textbook reference material rather than skill-specific guidance. | 2 / 5 |
Actionability | The tool guidance is fully concrete: exact tool names with parameter names in the tables ('species', 'brain_structure', 'dandiset_id'), a question-type-to-tool selection strategy, and arrow-chained call sequences for each resource (e.g. 'ModelDB_list_models -> pick a real ID -> ModelDB_get_model -> take the object_id from model_paper and call ModelDB_get_paper'). Verified live examples with real IDs and values (model 3263, ds000117, collection 457, neuron 147055) make the guidance copy-paste ready. | 5 / 5 |
Workflow Clarity | Sections 7-10 have clear multi-step sequences with explicit checkpoints and error-recovery guidance ('OpenNeuro_get_snapshot_validation ... to check BIDS-compliance first', the empty-edges-vs-count gotcha, the NeuroVault broken-search caveat), matching anchor 5 for those sections. It sits at 4 overall because the earlier reasoning-pattern workflows are weaker — the computational strategy says 'ALWAYS use Python' with no example code, and the clinical/anatomy patterns rely on 'VERIFY with literature' without a retry loop. | 4 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), so everything is inline in one 357-line file. The textbook material (sections 1-4) and the per-resource tool documentation clearly belong in separate reference files, but in-file structure is decent (numbered sections, per-resource workflow blocks, and a Sec. 11 cross-reference table for navigation). This is anchor 3's 'some structure but could be better organized; content that should be separate is inline' rather than anchor 2, since headers and navigation within the file are clear. | 3 / 5 |
Total | 14 / 20 Passed |