Content
60%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.
The body is a well-organized, unusually concrete orchestration spec with explicit pipelines, thresholds, and decision rules, though it is repetitive and its research-justification prose could be trimmed. Its critical defect is that every one of its ~16 reference files is missing from the bundle, making the advertised progressive-disclosure structure hollow.
Suggestions
Ship the referenced files (reference/*.md and _common/*.md) with the skill, or inline the essential content and delete the dangling links — as delivered, none of the ~16 referenced paths resolve.
Deduplicate collaboration/routing information: it currently appears in the CAPABILITIES_SUMMARY comment, Trigger Guidance, the Collaboration table, and the Reference Map; keep one canonical table and link to it.
Compress the research-justification prose in the Never list to one-line rules with a single pointer to reference/persona-bias-audit.md and _common/AI_PERSONA_RISKS.md, and add explicit validate-and-retry feedback loops to the CONJURE/EVOLVE pipelines.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and prescriptive rather than tutorial-like, but collaboration information is repeated four times (the CAPABILITIES_SUMMARY comment, Trigger Guidance, the Collaboration section, and the Reference Map), and the Never list carries extended research justifications ("caricature effect", "bias laundering" prose) that could be one line plus a pointer. This matches anchor 3 — mostly efficient but with unnecessary repetition that could be tightened — rather than 4, where only minor trimming would be needed. | 3 / 5 |
Actionability | For a process skill, guidance is highly concrete: exact thresholds (confidence 0.40 ask-first, ≥5% drift trigger, ≥15% divergence with n≥50), numeric decay rates, source-weight tables, commands (/Cast conjure), and the registry path .agents/personas/registry.yaml. It stops short of anchor 5 because the executable detail (schemas, payload shapes, adapter formats) is deferred to reference files that are not present in the bundle, leaving minor gaps in what can actually be executed from this file alone. | 4 / 5 |
Workflow Clarity | Every operating mode has an explicit pipeline (e.g. CONJURE: INPUT_ANALYSIS -> DATA_EXTRACTION -> PERSONA_SYNTHESIS -> VALIDATION -> REGISTRATION), backed by a phase table with required actions, key rules, and read targets, plus VALIDATION phases and ask-first gates for risky operations. Anchor 4 fits (clear sequence, most checkpoints present) because explicit validate-fix-retry feedback loops for error recovery are not spelled out, which anchor 5 requires. | 4 / 5 |
Progressive Disclosure | The in-body structure is well signaled (Reference Map, per-recipe "Read First" column, one-level references), but the bundle contains no reference/, references/, scripts/, or assets/ directories — all ~16 referenced files (reference/persona-model.md, reference/registry-spec.md, _common/AI_PERSONA_RISKS.md, etc.) are dangling, so the navigation leads nowhere and the actual detailed content is inaccessible. Structure without deliverable references lands at anchor 2 (minimal effective structure) rather than 3, since the split content simply does not exist in the bundle. | 2 / 5 |
Total | 13 / 20 Passed |