Content
78%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 well-architected skill body: the two-phase workflow is clearly gated, outputs are fully specified down to file paths, and progressive disclosure is near-textbook with a load-when table pointing to six real, substantive reference files. The main improvement opportunity is removing the duplication between Core Principles and the Constraints section, which restates the same rules almost verbatim.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean — terse tables, numbered steps, a compact ASCII phase diagram — and assumes Claude's competence (no explanation of what strangler fig is, no library tutorials). The deduction is for redundancy: Core Principles 2–4 ('Always cite evidence', 'Always research before recommending', 'Never assume') are restated nearly verbatim in the Constraints MUST DO / MUST NOT lists, and the workflow tree duplicates the numbered steps. Not a 3 because the padding is duplication of genuinely load-bearing rules rather than unnecessary explanation. | 4 / 5 |
Actionability | For an instruction-only skill the guidance is concrete: exact output paths ('./migration-plan/research/', './migration-plan/domains/', '00-roadmap.md'), named output files with their required contents, per-step reference loading, and a 'Cite every finding as file:line' rule that is directly checkable. Not a 5 because the actual methodology lives entirely in the reference files — appropriate for progressive disclosure, but the body alone gives shape without method (e.g., no example of a domain plan or risk-matrix entry). | 4 / 5 |
Workflow Clarity | Two explicitly gated phases ('Never skip RESEARCH. Never start PLAN without completing RESEARCH.') with eight numbered steps, per-phase outputs, and MUST DO/MUST NOT checklists acting as a verification list; 'stop and ask when uncertain' provides an error-recovery path. Not a 5 because there is no explicit validation checkpoint on the produced artifacts (e.g., 'verify every claim in a domain file has a file:line or URL before finishing') — though as a planning skill with no destructive or batch operations, the rubric's cap-at-3 rule does not apply. | 4 / 5 |
Progressive Disclosure | Exemplary structure: a reference table with a 'Load When' column, 'Load references based on the current phase and need. Do not preload all references.', all six referenced files exist and are substantive (181–236 lines each), and references are exactly one level deep with no nested chaining. The SKILL.md stays a true overview while methodology lives in the bundle. | 5 / 5 |
Total | 17 / 20 Passed |