Content
86%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 strong, dense skill document: executable code for both data paths, a clear six-step workflow with mandatory human-coder validation checkpoints, and a well-signaled one-level reference for chapter ranges. Minor weaknesses are slight repetition of the licensing/public-domain point and the absence of explicit error-recovery guidance for failed or empty searches.
Suggestions
State the licensing/public-domain constraint once (e.g. in 'Edge cases & gotchas') instead of three times across the intro, gotchas, and description.
Add a short error-recovery step to the workflow, e.g. what to do when a search returns no candidates (broaden the term, fall back to chapter-range browsing) or when the NLM API call fails (fall back to the local CMS table).
Trim the 'When to use' section, which largely restates the frontmatter description; a single pointer to the chapter reference would suffice.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Quotes: "ICD-10-CM and ICD-10-PCS are public domain. CMS publishes the complete annual code files..." and the intro paragraph "This is decision support for a certified coder, not autonomous billing". The body is dense and domain-specific (GEMs lossy, Excludes1/Excludes2, combination codes, Oct 1 fiscal-year updates — all non-obvious), but the public-domain/licensing point is stated three times (intro, gotchas, description) and the 'When to use' section repeats the frontmatter, so minor trimming is possible per the score-4 anchor. Not 5: there is a small amount of repetition; not 3: the padding is minor and most sections earn their tokens. | 4 / 5 |
Actionability | Quotes: "billable = line[14] == '1'" (concrete CMS file column parsing), "requests.get('https://clinicaltables.nlm.nih.gov/api/icd10cm/v3/search', params={'sf': 'code,name', ...})", and "openmed.analyze_text(note, model_name='disease_detection_superclinical', output_format='dict')". Two fully executable search paths plus an executable NER hand-off snippet, with copy-paste-ready code, concrete column offsets, API URL, FHIR system URI, and a named model — matching the score-5 anchor. Not 4: the common cases (local CMS file and public API search, plus entity consumption) are all covered with runnable code. | 5 / 5 |
Workflow Clarity | Quotes: the six-step Workflow ("Extract... Route to a chapter... Search... Apply ICD-10-CM specificity rules... Rank... Emit"), with validation checkpoints like "Flag where the note lacks the detail a billable code requires" and "marked status: needs-coder-review", plus "Human-in-the-loop is mandatory". The sequence is clear with most checkpoints present, but there is no explicit error-recovery guidance (e.g. what to do when a search returns no candidates or the API call fails), matching the score-4 anchor rather than the feedback-loop-rich score-5 anchor. | 4 / 5 |
Progressive Disclosure | Quotes: "see references/icd10-chapters.md for code ranges" and "Chapter/section ranges: references/icd10-chapters.md". The single referenced file exists, is one level deep, is clearly signaled with its purpose, and holds exactly the content that belongs outside SKILL.md (the 22-chapter range table), while the body stays a lean overview — matching the score-5 anchor. Not 4: there are no organization gaps; the split is appropriate. | 5 / 5 |
Total | 18 / 20 Passed |