Content
67%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 presents a well-sequenced iterative retrieval pattern with concrete relevance-scoring criteria, explicit stopping conditions, and illustrative worked examples. Its main weakness is that all code is pseudocode built on undefined helper functions, so a reader must invent the execution details. Structure and conciseness are solid.
Suggestions
Make at least the core loop executable: define or ground retrieveFiles, scoreRelevance, and refineQuery against real tooling (e.g., grep/glob commands or an actual function implementation) so the pattern can be applied directly.
Move the two worked examples and the relevance-scoring rubric into a references/ file and keep SKILL.md as a leaner overview, shortening the body toward the quick-start ideal.
Trim the ASCII diagram and the 'problem' section, which restate information already conveyed by the phase code and examples.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is largely efficient — problem/solution framing, compact code blocks, and two worked examples with no padding explaining concepts Claude already knows. Minor trimmable excess: the ASCII diagram and the 'problem' section partially restate what the code and examples already demonstrate. Not a 5 because every token does not quite earn its place; not a 3 because there is no significant over-explanation. | 4 / 5 |
Actionability | The guidance is structured around code (initialQuery, evaluateRelevance, refineQuery, iterativeRetrieve), but every snippet depends on undefined helpers (retrieveFiles, scoreRelevance, extractPatterns, mergeContext, hasCriticalGaps), making it effectively pseudocode — matching 'pseudocode instead of executable code; missing key details'. The agent-prompt block in 'エージェントとの統合' is the most directly usable piece, lifting it above 2. | 3 / 5 |
Workflow Clarity | The four phases (DISPATCH → EVALUATE → REFINE → LOOP) are clearly sequenced with an explicit stopping rule (>= 3 files at relevance >= 0.7, no critical gaps, max 3 cycles) — a real feedback loop with checkpoints. It falls short of 5 because the steps are presented as illustrative pattern code rather than concrete, run-ready commands with error-recovery handling. | 4 / 5 |
Progressive Disclosure | There are no bundle files (references/, scripts/, assets/ are absent), and the body is well-organized into clearly signaled sections (phases, examples, integration, best practices) with no nested references — the one-level '関連項目' links are clearly signaled. It is not 5 because at ~200 lines the worked examples and scoring criteria arguably belong in reference files to keep SKILL.md a leaner overview. | 4 / 5 |
Total | 15 / 20 Passed |