Content
77%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
Highly actionable and clearly sequenced, with concrete tools, formats, and error-recovery checkpoints. It is held back by heavy repetition that bloats the body and an overview that is too long given that detailed procedures already live in separate files.
Suggestions
De-duplicate the governing rules: state "one decision per turn / stop at the question", "never invent a signal", "read-only / discovery-only", and "pass by trait_hash" once each and reference them, rather than restating them in Purpose, each flow step, and How to behave.
Move the per-step narration/format detail (candidate card columns, option label grammar, profile-view axes) into the existing context files so SKILL.md reads as a lean overview that links out, strengthening both conciseness and progressive disclosure.
Ensure the referenced context/discovery.md, context/scoring.md, and context/adjacencies.md are present in the bundle so the on-demand reads resolve, since no bundle directory currently contains them.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is ~22 KB / ~150 lines and restates its governing principles many times ("one decision per turn / stop" in the Purpose, each flow beat, and How to behave; "never invent a signal", "read-only", and "pass by trait_hash" each repeated across sections), so it is accurate but padded and could be tightened substantially. It is not a 1 because it does not pad with concepts Claude already knows. | 2 / 3 |
Actionability | Guidance is concrete and executable: named tool calls (trait_search, visualize, cat "<Bundle root>/context/discovery.md"), specific option formats ("name · ~size · fresh/standard"), numeric bounds (1–3 searches, 2–4 options, ≤2-line reads), named render types, and quoted refusal/example utterances. | 3 / 3 |
Workflow Clarity | A clearly sequenced 7-step flow with a per-beat stop checkpoint, a tracked kept-set/pool-view update on every change, and explicit feedback loops in Failure modes plus the first-call liveness gate for error recovery. | 3 / 3 |
Progressive Disclosure | The three heavy procedures are split into one-level-deep, clearly signaled context/*.md references ("read it once this session with cat ..."), but those files are not present in any bundle directory so cannot be verified, and the SKILL.md overview is itself a dense monolith with step detail that could live in those separate files. | 2 / 3 |
Total | 10 / 12 Passed |