Content
81%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 highly actionable, well-sequenced orchestration skill with concrete MCP prompts, explicit validation gates, and a complete output template. Its weaknesses are repetition of the same judgment principles across three phases and Key Rules, and references to shared files that are not part of the bundle.
Suggestions
Consolidate the repeated brainstorming-quality principles (named-risk-over-hedging, "apply X to Y" legitimacy, annotate-don't-eliminate) into a single Key Rules entry and reference it from Phases 2-4 instead of restating them in each prompt block.
Trim Phase 3's long blockquote rationale about cross-model reviewer design to 2-3 sentences — the operational rule (cluster duplicates, gate only on budget facts, annotate the rest) carries the instruction.
Bundle the shared output protocols (output-versioning.md, output-manifest.md, output-language.md) under references/ within the skill, or summarize their operative rules inline, so the skill does not depend on files outside its directory.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient prescriptive content Claude could not infer (MCP tool names, budget constants, phase structure), but the same principles are restated in multiple places — "annotate, do not eliminate" and the objective-gate rationale in Phase 3's blockquote and steps, the named-risk-over-hedging and "apply X to Y" guidance appearing in the Phase 2 prompt, the Phase 4 prompt, and Key Rules. This repetition could be consolidated. | 3 / 5 |
Actionability | Fully concrete: exact MCP tool calls with complete prompt blocks (mcp__gemini-review__review_start, review_reply_start, review_status), jobId save-and-poll protocol, named constants with override syntax, concrete file paths (idea-stage/IDEA_REPORT.md), and a copy-paste report template. Placeholders like [paste landscape map] are appropriate prompt-template slots. | 5 / 5 |
Workflow Clarity | Six phases are clearly sequenced with explicit validation checkpoints: GPU-hour estimates required BEFORE launching, hard timeout kill with partial-result collection, decision criteria defined upfront (including what null/negative outcomes teach), budget tracking against MAX_TOTAL_GPU_HOURS, and an evidence-based re-ranking loop that feeds back into the final report. | 5 / 5 |
Progressive Disclosure | Good structure with well-signaled one-level-deep references in a dedicated Output Protocols section, and appropriately inlined material (the report template is used directly). Not 5: the shared-protocol links (../../shared-references/output-versioning.md etc.) resolve outside this bundle, which contains no references/, scripts/, or assets/ directories, so navigation depends on files not shipped with the skill. | 4 / 5 |
Total | 17 / 20 Passed |