Content
62%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 delivers an exceptionally well-sequenced workflow with genuine validation checkpoints, budgets, timeouts, and fallback paths on every risky operation, and its MCP/bash guidance is largely copy-paste executable. Its weaknesses are redundancy and monolithic structure: the reviewer-routing contract is repeated across six locations and substantial script/template content is inlined rather than split into reference files, inflating token cost for every invocation.
Suggestions
State the reviewer-backend calling convention once (e.g. in a dedicated section or shared reference) and have the header blockquote, Constants, and Phases 2/4 link to it instead of restating it.
Move the ~55-line Phase 0 wiki/threat-scanner resolver bash into a scripts/ file and invoke it, keeping only the fail-closed usage rules inline.
Extract the full IDEA_REPORT.md template and the brainstorm/triage bundle prompts into references/ files to cut the body's size and per-invocation token cost.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The reviewer-backend routing contract is restated roughly six times (header blockquote, Constants, Reviewer Calling Convention, Phases 2 and 4, Review Tracing), the "never drop a candidate for being weak" rule appears three times, and a ~55-line bash resolver is inlined in Phase 0 — several padded/redundant sections fitting anchor 2. It avoids anchor 1 because it never explains concepts Claude already knows; the density is duplication, not tutorial filler, but it is more than the "some" tightening of anchor 3. | 2 / 5 |
Actionability | Provides mostly executable guidance: a robust copy-paste bash resolver with fallback chains, exact MCP call configs (model, reasoning effort, threadId reuse), JSON output schemas, and a complete report template. Minor gaps — the Phase 7 block is pseudo-shell ("if research-wiki/ exists AND [ -n ... ]:") rather than runnable code — keep it at anchor 4 instead of the fully-executable anchor 5. | 4 / 5 |
Workflow Clarity | Phases 0-7 are clearly sequenced with explicit validation checkpoints and error-recovery loops throughout: a fail-closed threat scan before reading the query pack, upfront pilot budgets with hard timeouts (PILOT_MAX_HOURS, MAX_TOTAL_GPU_HOURS), decision criteria defined before launch, and WARN-and-continue fallbacks on every failure path. This matches the anchor-5 validate-fix-retry pattern. | 5 / 5 |
Progressive Disclosure | References to ../shared-references/*.md are clearly signaled and one level deep, but the 612-line body is near-monolithic: the bash resolver script, the full report template, and repeated routing rules are inlined where they belong in scripts/ or references/, and the skill ships no bundle files of its own. Structure exists and references are well signaled, so this fits anchor 3 rather than 2, but the scale of inlining rules out anchor 4. | 3 / 5 |
Total | 14 / 20 Passed |