Content
53%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 domain half of this skill is excellent — verified, copy-paste-ready code against a real script with well-signaled one-level-deep references. The body is dragged down by a large tail of generic, non-domain agent-contract boilerplate, a nonsensical audit command, an irrelevant study-design checklist, and an abstract workflow that never sequences the actual cleaning pipeline.
Suggestions
Delete the ~80 lines of generic agent-contract boilerplate (Output Requirements, Error Handling, Input Validation, Response Template, When Not to Use, Required Inputs, Output Contract, Failure Handling, User Checkpoints, Quick Validation) and the duplicated Quick Check/Audit-Ready Commands section — they are non-domain, largely duplicate each other, and consume context without teaching the cleaning task.
Rewrite the Workflow section as the concrete cleaning pipeline with validation checkpoints, e.g. load → `validate_domain()` (stop and report missing required fields) → `handle_missing_values()` → `detect_outliers()` → `standardize_dates()` → `save_report()`, with a concrete validation command (e.g. re-run `validate_domain` on cleaned output) before declaring completion.
Remove the broken audit command (`--input "Audit validation sample with explicit symptoms..."` passes prose as a file path) and the Pre-Cleaning checklist items about IACUC approval, statistical power, and randomization, which are study-design concerns outside this skill's data-cleaning scope.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The core capability sections are lean and code-first, but the body carries substantial padding: ~80 lines of generic agent-contract boilerplate ("Output Requirements", "Error Handling", "Input Validation", "Response Template", "When Not to Use", "Required Inputs", "Output Contract", "Failure Handling", "User Checkpoints", "Quick Validation" — three of which duplicate each other), a "When to Use" bullet that mechanically repeats the frontmatter description inside a broken sentence ("Use this skill when the task needs Use when cleaning clinical trial data..."), a duplicated py_compile command in both "Quick Check" and "Audit-Ready Commands", and an irrelevant study-design checklist ("IACUC approval obtained", "Sample size adequately powered", "Randomization method documented") that is not data cleaning. This matches anchor 2: "Noticeably verbose; several unnecessary explanations or padded sections". | 2 / 5 |
Actionability | The Quick Start, per-capability snippets, Complete Pipeline, CLI examples, threshold table, and required-field lists are concrete and were verified executable against scripts/main.py (class name, all methods, and CLI flags match). Not 5 because one example is broken as written — `python scripts/main.py --input "Audit validation sample with explicit symptoms, history, assessment, and next-step plan."` passes a prose string as an input file path — a minor gap consistent with anchor 4. | 4 / 5 |
Workflow Clarity | The "Workflow" section lists six steps but they are a generic consulting template ("Confirm the user objective", "Return a structured result") rather than the actual cleaning pipeline (load → validate_domain → handle_missing_values → detect_outliers → standardize_dates → save_report, which only appears implicitly as the capabilities list). The skill batch-processes datasets and overwrites output files, yet validation is abstract prose with no concrete in-flow checkpoint, so per the rubric cap workflow_clarity cannot exceed 3. | 3 / 5 |
Progressive Disclosure | Good structure: SKILL.md is an overview with quick-start and capability summaries, and details are split into five real, verified, one-level-deep reference files (common-patterns.md, troubleshooting.md, sdtm_ig_guide.md, domain_specs.json, outlier_thresholds.json) plus a working script, each clearly signaled with links and content summaries. Not 5 because the inlined generic contract sections are content that belongs in no file at all, and the 262-line body is heavier than a clean overview should be. | 4 / 5 |
Total | 13 / 20 Passed |