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-structured advisory skill with unusually concrete mode detection, workflow formats, and routing, plus dense non-obvious principles. Its main defect is that the progressive-disclosure architecture points at bundle files (rules/ and templates/) that are not present, and a secondary one is a duplicated rule-file table costing ~20 lines.
Suggestions
Include the referenced 'rules/*.md' and 'templates/*.md' files in the skill bundle (or fix the paths) — 17 of the 19 referenced paths do not resolve, which breaks the on-demand loading design the skill depends on.
Merge the 'Area Routing' and 'Required Reading by Area' tables: they duplicate the same 13 rule-file links, so a single table with load conditions plus a short list of references/templates would save ~20 lines.
Add an inline validation checkpoint to the 'review' workflow — e.g., confirm each detected area maps to a rule file that was actually loaded before emitting findings — so verification is not only end-of-process via the Definition of Done.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is a lean index that assumes Claude's competence — Core Principles and the Anti-patterns one-liners are dense, non-obvious content with no padding. Not 5: the 'Required Reading by Area' table repeats the same 13 rule-file links already given in the Area Routing table (~20 trimmable lines). Not 3: beyond that duplication, every section earns its place and nothing explains concepts Claude already knows. | 4 / 5 |
Actionability | For an instruction-only skill the guidance is fully executable: exact mode-detection conditions ('$0 == "review"', 'a file/path is supplied as $ARGUMENTS'), a literal output block ('Mode: review / Areas: ... / Targets: src/agents/triage.ts (system prompt at L24-78)'), a concrete findings format (What/Rule/Fix with 'path:line' and a 'Top 3 fixes' list), a single batched question list for design mode, explicit template mapping, and exact Skill() invocation calls with a documented fallback. Per the scoring notes, absence of code in an instruction-only skill is not penalized when the guidance is this actionable. | 5 / 5 |
Workflow Clarity | Three clearly sequenced numbered workflows with real checkpoints: state mode/areas before continuing, 'Do not edit the file in review mode unless the user asks', a Definition of Done checklist, a clarifying-question rule when no area is named, and a fallback when OTEL skills are unavailable. No destructive or batch operations, so the validation cap does not apply. Not 5: validation is mostly end-of-process (the DoD checklist) rather than inline per-step feedback loops — e.g., no step verifies that each detected area actually maps to a loaded rule file before findings are produced. Not 3: checkpoints are explicit, not implicit, and error paths are handled. | 4 / 5 |
Progressive Disclosure | The design intent is exemplary — a self-described 'thin index' with one-level-deep, well-signaled references and per-area load conditions — but scored against the actual bundle, 17 of the 19 referenced paths (all 13 'rules/*.md' and all 4 'templates/*.md') do not exist; only 'references/primary-sources.md' and 'references/recent-changes.md' are present. Not 4: the gap is not minor organization — the majority of referenced files are absent, so the on-demand loading chain cannot actually be followed. Not 2: the structure itself is good, references are prominently signaled rather than buried, and nothing that belongs in separate files is inlined. | 3 / 5 |
Total | 16 / 20 Passed |