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.
A superbly written control document — lean, well-sequenced, with a thoughtful mode-dispatch and on-demand reading table — that is undermined by its bundle: only references/frameworks.md ships, while all rules/ and templates/ files the workflow depends on are absent. As delivered, the skill can route and frame a request but cannot execute its diagnosis, ethics gate, or emission steps.
Suggestions
Ship the referenced bundle files (rules/diagnosis.md, rules/interventions.md, rules/dual-process.md, rules/ethics.md, rules/workshop.md, templates/intervention-plan.md, templates/workshop.md) or inline their essential content into SKILL.md — currently 7 of 8 referenced paths are dangling, which is the single largest quality defect.
Inline the three-question ethics test (or its essence) in Step 4 so the mandatory gate remains executable even if rules/ethics.md fails to load, mirroring the self-contained fallback pattern already used for missing composed skills.
Add a one-line fallback for missing rule files, e.g. proceed with the Hard Rules as the minimum viable rule set and report which files were skipped, matching the existing "A missing skill never blocks" convention.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and assumes Claude's competence: frameworks are named ("COM-B, Fogg B=MAP") without tutorials, the ✗/✓ rewrite example in Step 1 teaches the rule in two lines, and the mode and reading tables replace prose entirely. No section explains concepts Claude already knows, so nothing is trimmable without losing function — anchor 4's "minor instances of over-explanation" does not apply. | 5 / 5 |
Actionability | The inline process scaffolding is concrete (the who/what/when/measured rewrite formula with a worked example, the batched four-item context checklist with `unknown` handling, testable Hard Rules like "Every intervention has a metric and a baseline-or-`unknown`"), but the operational core is deferred to files absent from the bundle: Step 3 says "Follow the loaded rule files", Step 4 gates output on "the three-question test in rules/ethics.md", and Step 5 fills "the mode's template" — and of the eight referenced paths only references/frameworks.md exists. This matches the concrete-but-incomplete anchor; it is not 4 because the gaps are the levers, ethics test, and templates themselves. | 3 / 5 |
Workflow Clarity | Steps 1–5 are clearly sequenced with real checkpoints: the attitude-vs-behavior check in Step 1 ("ask for the action once"), unknown-baseline recovery in Step 2 ("proceed with `unknown`s… never stall"), the Step 4 ethics gate ("Any intervention that fails is dropped from the output and replaced by its honest alternative"), and the Hard Rules acting as an emission checklist. It is not 5 because the central validation checkpoint's content — the three-question ethics test — lives only in the missing rules/ethics.md, so the gate is named but not executable as written. | 4 / 5 |
Progressive Disclosure | Scored against the actual bundle structure: the body designs an exemplary one-level-deep, load-on-demand scheme (the "Required Reading by Mode" table, "Load on demand — never all up-front", and references/frameworks.md § Evidence grading — which does exist), but only 1 of the 8 referenced paths is present; the five rules/*.md and two templates/*.md files are missing, so navigation to the core content is impossible in practice. The majority of paths being dangling pulls this between anchor 2 and anchor 3, landing at 2. | 2 / 5 |
Total | 14 / 20 Passed |