Content
73%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, highly actionable gate skill with an exemplary decision workflow — explicit sequence, validation checklist, and error-recovery branches throughout. Its weaknesses are redundancy (the anti-hardcoding rule repeated ~5×, an inlined book-category list that contradicts it) and a lookup example that only demonstrates the entity primitive, leaving the reader to adapt it for the other match types.
Suggestions
State the 'reads the ACTIVE schema pack, never a hardcoded table' rule once (e.g., in the Critical section) and have the Contract, Step 5, Anti-Patterns, and Hard Rules sections reference it instead of restating it — this would cut substantial tokens.
Drop the inlined 11-item book-category list from Step 3; replace it with a jq command that enumerates the active pack's media/books path prefixes, which both shrinks the skill and removes the hardcoding the skill itself forbids.
Generalize the Step 2 example to show how to look up any primitive (entity|temporal|concept|media|annotation) — e.g., a parameterized jq filter or one command per primitive — so all seven Step 1 outcomes have an executable lookup.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The 'no hardcoded directory table' rule is restated roughly five times across the Contract, the 'Critical' section, Step 5, Anti-Patterns, Hard Rules, and the Changelog, and Step 3 inlines an 11-item book-category list immediately before telling the reader to walk the pack instead of hardcoding it. The Changelog also carries version-sensitive references ('v0.39+ broke the old hardcoded table', 'D9 from the v0.39 plan-eng-review') outside any deprecated section, which the rubric penalizes. Mostly efficient, domain-specific content that still needs tightening — anchor 3, not 2 (nothing explains generic concepts Claude already knows) and not 4 (the repetition is substantive, not minor). | 3 / 5 |
Actionability | Concrete, executable commands appear throughout (`gbrain schema show --json | jq '.page_types[] | select(.primitive == "entity")'`, `gbrain schema review-orphans --json`, `gbrain schema lint --with-db --json`, `gbrain schema add-type <type> --primitive <p> --prefix <dir/>`), plus a 7-question ordered decision protocol, a validation checklist, and copy-paste output templates. Anchor 4 rather than 5 because the Step 2 lookup example filters only the entity primitive while the protocol's other matches are temporal/media/source primitives, and delegation targets ('consult EIIRP Phase 3', 'route to the schema-author skill') are named without enough detail to execute. | 4 / 5 |
Workflow Clarity | The Decision Protocol is a clear 5-step sequence ending in an explicit 'Validate before writing' checklist, with error-recovery branches (no matching type → 'DON'T pick the closest-fitting one', signal EIIRP Phase 3) and drift thresholds ('When more than 10% of a source is untyped, run the EIIRP Phase 3 SCHEMA CHECK flow'). Validation checkpoints, feedback loops, and a dedicated 'When it fails' routing section are all present — anchor 5; the destructive-operation cap does not apply since the skill is explicitly advisory (`mutating: false`). | 5 / 5 |
Progressive Disclosure | No bundle files exist, so all content sits in one well-sectioned ~190-line SKILL.md with clear headers (Purpose, Contract, Decision Protocol, Drift Detection, Output Format, Anti-Patterns), which is a reasonable single-file size; the only external pointer ('../conventions/agent-operator-protocol.md') is one level deep. Good structure matching anchor 4, not 5 because the file exceeds the simple-skill threshold and inlines filler (Changelog, full book-category list) that adds no navigational value. | 4 / 5 |
Total | 16 / 20 Passed |