Content
56%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 technical core is strong — concrete commands, a real script with matching classes, and detailed input/output schemas with a genuine validation loop. But nearly half the file is generic template boilerplate unrelated to this skill, the Parameters table is corrupted with duplicated rows, and Prerequisites references a nonexistent requirements.txt, diluting token efficiency and leaving the file at the rubric midpoint.
Suggestions
Cut the ~150 lines of generic boilerplate (Security Checklist, Response Template, Output Contract, Failure Handling, Required Inputs, User Checkpoints, Input Validation, Output Requirements, Quick Validation) down to the single most useful version, or move them to a shared reference file — most restate the same assumptions/partial-completion guidance.
Fix the factual errors: remove the nonexistent 'pip install -r requirements.txt' (the skill is pure Python with no dependencies), repair the Parameters table's duplicated --input/--output rows and the empty Description for --phase, and correct the 'from scripts.main import ...' path.
Rewrite the five-step 'Workflow' to follow the actual tool sequence (generate -> optimize -> analyze -> benchmark) with the validation checkpoints embedded as steps, and move the criteria-categories and optimization-strategies tables into the existing reference files with short pointers.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Roughly 150 of the ~350 lines are generic governance boilerplate — 'Security Checklist' (13 unchecked items about APIs, HTTPS, and sandboxing for a pure-Python local tool), 'Output Requirements', 'Input Validation', 'Response Template', 'When Not to Use', 'Required Inputs', 'Output Contract', 'Failure Handling', 'User Checkpoints', and 'Quick Validation' — that repeatedly restate the same assumptions/partial-completion guidance in different words. 'Quick Check' also duplicates the py_compile command from 'Audit-Ready Commands', and the Parameters table repeats --output four times and --input three times. This is noticeably verbose (anchor 2); it is not a 1 because the skill-specific sections (CLI usage, JSON formats, categories, strategies) are reasonably tight and do not explain concepts Claude already knows. | 2 / 5 |
Actionability | The body provides concrete, verified executable guidance: copy-paste CLI invocations for all four subcommands (generate/optimize/analyze/benchmark), a Python API example whose classes CriteriaGenerator and CriteriaOptimizer exist in scripts/main.py, and full input/output JSON schemas with realistic clinical examples. It is not a 5: Prerequisites says 'pip install -r requirements.txt' but requirements.txt does not exist (contradicting 'Dependencies: None (pure Python)'), 'from scripts.main import ...' is not an importable package path, and the Parameters table has duplicate rows and an empty Description cell for --phase. | 4 / 5 |
Workflow Clarity | There is a genuine feedback loop: 'Quick Check' (python -m py_compile scripts/main.py) verifies the script parses before deeper execution, 'Audit-Ready Commands' validate the CLI, and 'Error Handling' plus 'Failure Handling' specify reporting the failure point and a manual fallback. It is not a 5 because the five-step 'Workflow' itself is abstract governance language ('Validate that the request matches the documented scope', 'Return a structured result') rather than the concrete generate -> optimize -> analyze -> benchmark tool sequence, so checkpoints sit beside the workflow instead of inside it. | 4 / 5 |
Progressive Disclosure | Scored against the actual bundle: the five listed reference files (criteria_templates.json, optimization_guidelines.md, common_pitfalls.md, regulatory_guidance.md, feasibility_data.json) all exist, and scripts/main.py exists with the documented classes and subcommands — references are real and one level deep. However, substantial content that belongs in those reference files or a shared template is inlined in SKILL.md: the criteria-categories table, optimization-strategies tables, parameter tables, and the ~150 lines of generic boilerplate sections. Anchor 3 ('Some structure but could be better organized... content that should be separate is inline') fits; it is not a 4 because that inlining is significant, and not a 2 because the references are present, clearly listed, and loadable rather than buried. | 3 / 5 |
Total | 13 / 20 Passed |