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 is a well-architected operational contract with clear phase sequencing, explicit routing tables, and genuinely actionable dispatch behavior for each recipe. Its main costs are repetition of the cache policy across three sections and metadata duplicated between the HTML comment header and the Collaboration section, plus some procedural detail that belongs in its referenced files.
Suggestions
State the cache-first policy once (in the Core Contract) and have the CACHE workflow row and the `recommend` dispatch note point back to it instead of restating the rules, cutting three overlapping passages to one.
Drop the duplicated collaboration metadata from the opening HTML comment block (BIDIRECTIONAL_PARTNERS / COLLABORATION_PATTERNS) since the Collaboration section already carries the same partner and handoff information in table form.
Keep only a one-line summary of the init/refresh flow in the body and move the full SCAN→SIZE→SCORE→PICK→WRITE→REPORT detail (signal sources, H/M/L scoring, top_n sizing) into `reference/cache-recipes.md`, which the body already designates for it.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is information-dense with no explanations of concepts Claude already knows, but it could be tightened: the cache-first policy is restated in the Core Contract, the CACHE workflow row, and the `recommend` dispatch notes, and the opening HTML comment block duplicates the Collaboration section's partner/handoff metadata. This repetition is noticeable but structural rather than padding, sitting above the verbose level of 2. | 3 / 5 |
Actionability | Guidance is largely executable for an instruction-only skill: concrete subcommand dispatch rules, named signal sources (package.json, Cargo.toml, file-extension distribution), a sized top_n range (15-50), and exact prompt text ("Generate cache? (Y/n) — reduces context ~95%..."). It stops short of 5 because the actual recommendation/comparison output formats are deferred to reference files rather than exemplified inline. | 4 / 5 |
Workflow Clarity | The LISTEN → CACHE → MATCH → RECOMMEND → ORIENT table clearly sequences phases with per-phase files to read, and checkpoints exist for fragile spots (catalog_version/TTL staleness warning, ask-before-overwriting an existing cache, confirm-narrowing when exceeding the cap of 3). It lacks the explicit validate→fix→retry feedback loops of the 5 anchor, though the navigation task is not destructive or batch-oriented. | 4 / 5 |
Progressive Disclosure | Structure is strong: a Reference Map with "Read this when..." guidance, a per-recipe "Read First" column, and one-level-deep references to reference/*.md and _common/*.md. Minor gaps keep it below 5: the full SCAN→SIZE→SCORE→PICK→WRITE→REPORT procedure is inlined in the body despite `reference/cache-recipes.md` existing for it, and the _common spine-contract list adds a second layer of indirection; no bundle files were present to verify the referenced paths. | 4 / 5 |
Total | 15 / 20 Passed |