Content
77%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.
An exceptionally actionable and well-sequenced workflow with strong validation checkpoints, state persistence, and honest anti-fabrication rules. Its weaknesses are length and self-repetition (agency norms stated in three places) plus a monolithic structure that inlines eight agency specifications and verbose MCP prompt blocks that belong in separate reference files.
Suggestions
Move the eight Grant Type Specifications tables into one-level-deep reference files (e.g., references/kakenhi.md, references/nsfc.md, references/erc.md) and keep only a compact selection table in SKILL.md, which would address the weakest dimensions (progressive_disclosure and conciseness) at once.
Deduplicate agency norms: they currently appear in Grant Type Specifications, again in Grant-Specific Drafting Guidelines, and again in Key Rules ("Cultural norms matter"); state each norm once and cross-reference it.
Extract the Round 1 / Round 2+ MCP prompt templates and the repeated checkpoint option lists into a single shared pattern or a references/review-protocol.md instead of restating near-identical boilerplate at every phase.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most content is domain knowledge Claude does not have (agency sections, sub-types, review criteria), so it avoids the known-concept padding of anchor 2, but there is real redundancy: agency cultural norms appear three times (Grant Type Specifications tables, Grant-Specific Drafting Guidelines, Key Rules), the "grant argues for future work, not completed work" point is stated three times, the pipeline is diagrammed twice, and the checkpoint option lists are near-identical boilerplate repeated at every phase. It sits squarely at "mostly efficient but could be tightened". | 3 / 5 |
Actionability | Guidance is fully executable: exact input/output file paths (grant-proposal/GRANT_PROPOSAL.md, GRANT_STATE.json with a concrete JSON schema), literal sub-skill invocation commands (/research-lit "$ARGUMENTS", /novelty-check "[proposed gap statement]"), named MCP tools (mcp__gemini-review__review_start, review_reply_start, review_status) with prompt templates, a specific drafting order, and a ten-item final checklist covering the common cases. | 5 / 5 |
Workflow Clarity | Phases 0–5 are explicitly sequenced with numbered steps, each phase has an explicit validation checkpoint ("⛔ STOP HERE and wait for user response") with enumerated user options and fallback behavior, the external-review phase includes a review→revise→re-review feedback loop bounded by MAX_REVIEW_ROUNDS with an MCP-unavailable fallback, and GRANT_STATE.json enables recovery after compaction — matching the anchor-5 pattern of clear sequence, explicit validation, and error-recovery loops. | 5 / 5 |
Progressive Disclosure | There are no references/, scripts/, or assets/ directories at all — everything lives inline in a ~630-line SKILL.md. The eight agency specification tables (~90 lines) and the full MCP prompt blocks for the review rounds are content that clearly belongs in one-level-deep reference files (e.g., references/kakenhi.md, references/nsfc.md). Section headers make it navigable and the shared-references links are clearly signaled, but content that should be separate is inline, which is the anchor-3 fit; it is not anchor 2 because the document is well-sectioned and not a monolithic wall. | 3 / 5 |
Total | 16 / 20 Passed |