Content
71%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-structured orchestrator skill body: precise routing rules, a sensible precedence model for repo-vs-skill conflicts, and a clear tiered reporting workflow with user-approval gates on mutation. Its main defect is bundle integrity — the six agent files the whole workflow depends on are referenced but absent, which undermines both navigation and executability.
Suggestions
Ship the six agents/*.agent.md files (typescript, react, csharp, markdown, intent, constitution) in the bundle, or remove the agent-mode routing and inline the per-language checks — currently Step 1 routes every review to files that do not exist.
Trim the 11-bullet trigger list in 'When to use' and the 'When not to use' section, which restate the description's USE FOR / DO NOT USE FOR content and add token cost without new information.
Add a post-correction verification step (e.g., re-run the convention check on edited files and confirm no new Required findings) to close the workflow feedback loop after applying user-approved corrections.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is largely lean — routing table, precedence rules, and the four-step workflow all assume Claude's competence with no concept explanations. The 11-item trigger list and the 'When not to use' section partially duplicate the description's USE FOR / DO NOT USE FOR clauses and could be trimmed. | 4 / 5 |
Actionability | Guidance is concrete and executable for an instruction-only skill: file-extension classification rules ('.ts, non-component .tsx'), a named agent routing table, tiered finding categories (Required / Recommended / Advisory), and explicit correction policy. Minor gaps only — e.g., no example of a finding's 'state the rule, affected code, corrected version' format. | 4 / 5 |
Workflow Clarity | The four steps (classify/route, apply, present, correct) are clearly sequenced with explicit checkpoints — 'If language cannot be determined, ask before proceeding' and 'Apply only corrections the user explicitly approves'. A post-correction re-check loop is absent, keeping it below a 5. | 4 / 5 |
Progressive Disclosure | Section structure and reference signaling are good (agent-mode table, one-level-deep references/*.conventions.md, assets/*.follow-up.md that all resolve), but scored against the actual bundle: the six agents/*.agent.md files that Step 1 routes every review through do not exist in the bundle, leaving the primary execution path pointing at missing files. This is more than a minor organization gap, so it sits below 4. | 3 / 5 |
Total | 15 / 20 Passed |