Content
75%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 instruction skill: explicit phases, constants, decision tables, and refusal handling, with no padding and a good failure-mode philosophy (guide rather than error). Main weaknesses are the unresolved external reference paths and a few unspecified input-parsing mechanics (state file, buffer computation).
Suggestions
Resolve or inline the ../../shared-references/ links (cadence-protocol.md, candidate-schema.md) — they leave the skill bundle, so the skill cannot stand alone; either ship them as references/ files or summarize the needed rules (buffer color thresholds, candidate entry schema) inline.
Add a Phase 2 fallback for missing inputs: what to do when predictions/*.md or .cheat-state.json does not exist (e.g., treat published IDs as empty, assume green buffer) so the dedup and buffer-override steps cannot fail silently.
Tighten the Phase 0 onboarding script and Phase 4 sample outputs into a short template plus one condensed example — the two verbatim blocks are the largest token cost and partially duplicate the Key Rules.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence (composite, tier, buffer colors are never explained, only used), with no filler — but the ~30-line verbatim onboarding script in Phase 0 and the long sample output blocks could be tightened. Fits 'efficient; minor instances of over-explanation that could be trimmed'; not 5 because those template blocks earn their tokens only partially. | 4 / 5 |
Actionability | Exact constants (TOP_N=5, lookback formula), enumerated filter rules, a buffer-color decision table, verbatim output templates, and explicit refusal handling make this mostly executable instruction guidance; minor gaps remain — how to read state.shoots/.cheat-state.json and compute buffer color is delegated out, and predictions/*.md header format is assumed. Not 5 for those unspecified mechanics, not 3 because the guidance is concrete, not pseudocode. | 4 / 5 |
Workflow Clarity | Phases 0–4 are clearly sequenced with a flow diagram, an existence checkpoint in Phase 0, and a validation-with-recovery loop in Phase 1 (malformed entries → ask the user, never silently ignore), plus decision tables for buffer overrides. Not 5 because Phase 2 has no fallback if predictions/*.md or .cheat-state.json is absent, which would break the dedup/state steps; not 3 because checkpoints are explicit and recovery paths exist. | 4 / 5 |
Progressive Disclosure | Well-organized sections (Overview/Constants/Inputs/Workflow/Key Rules/Refusals/Integration) with clearly signaled one-level-deep links ([cadence-protocol.md], [candidate-schema.md]). Not 5 because both referenced paths point outside the skill bundle (../../shared-references/) and no references/ directory ships with the skill, so the links cannot be resolved from the bundle itself — a structural, if minor, navigation gap. | 4 / 5 |
Total | 16 / 20 Passed |