Content
82%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 strong, dense body: fully executable quick start, a concrete 5-step workflow with real cue lists, thoughtful hand-off contracts to sibling skills and FHIR, and honest scope/edge-case treatment. Minor gains available in trimming the motivational intro and citations, adding an explicit re-validation pass for spans left unclassified, and potentially splitting cue-lexicon detail into a reference file.
Suggestions
Trim the six-line introductory paragraph to a single sentence ('NER finds mentions; it does not tell you whether the patient has them — the ConText layer assigns negation, temporality, and uncertainty before grounding'), and consider condensing the Standards & references section to bare citation names with DOIs on one line each.
Add an explicit validation step to the workflow (e.g., 'Step 6 — Verify: any span that resolves to AFFIRMED/RECENT/CERTAIN with no cue analysis, or that came back with an unexpected axis, re-inspect the window and cue list before recording'), creating a feedback loop rather than relying on the edge-case note to 'double-check anyway'.
Consider moving the per-axis cue lexicons and the full function/constants reference into a one-level-deep references/ file (e.g. CUES.md), keeping SKILL.md to the workflow and quick start — the body currently inlines API-reference material that would scale poorly as the cue lists grow.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is efficient and teaches only non-obvious domain specifics (cue lexicons, pseudo-negation masking, even/odd parity, FHIR status mapping) rather than concepts Claude already knows. However, the six-line motivational intro ('NER finds *that* a condition was mentioned...') could be one line, and the 'Standards & references' citation block is padding-adjacent, so it fits 'efficient; minor instances of over-explanation that could be trimmed' rather than the every-token-earns-its-place anchor. | 4 / 5 |
Actionability | The quick start is copy-paste runnable — real imports (openmed.clinical), a concrete note string, and annotated expected output per span ('chest pain -> negated recent certain (do NOT record as present)') — and the workflow steps give exact function signatures with the actual cue lists ('denies', 'no evidence of', 'history of', 's/p', 'rule out', 'probable', 'vs'). This matches 'fully executable; copy-paste ready code; specific examples cover the common cases'. | 5 / 5 |
Workflow Clarity | The 5-step workflow is clearly sequenced with per-step function calls, return values, and downstream application rules, and the quick-start comments act as expected-output checkpoints. It sits at 'clear sequence with most checkpoints present; minor validation gaps' rather than 5 because there is no explicit verify/re-check loop (e.g., re-inspect spans the cue lists leave unclassified) beyond the edge-case advice to 'double-check anyway'. | 4 / 5 |
Progressive Disclosure | The skill is self-contained with no bundle files, and the sections (When to use, Quick start, Workflow, Hand-off, Edge cases, Standards) are well-organized and flat — no nested or buried references. It scores 4 rather than 5 because, at ~118 lines, the per-axis cue lexicons and function reference could plausibly live in a one-level-deep reference file per the anchor-5 pattern, though inlining at this length is defensible. | 4 / 5 |
Total | 17 / 20 Passed |