Content
73%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 an unusually well-engineered multi-phase workflow with explicit entry/exit criteria, layered validation with feedback loops, and rich agency-specific operational detail. Its main weaknesses are avoidable redundancy that inflates token cost, and phantom references to a templates/ directory, config.yaml, and CREDITS.md that do not exist in the bundle, which undercut both actionability and navigation.
Suggestions
Ship the referenced bundle files or remove the references: templates/us/ (nih_specific_aims.md, nsf_project_summary.md, budget_justification.md), templates/cn/ (nsfc_justification.md, nsfc_research_content.md, nsfc_research_foundation.md, nsfc_abstract_5sentence.md), config.yaml, and CREDITS.md are all cited in the body but missing from the bundle — every 'load the template' instruction currently dead-ends.
Deduplicate repeated guidance to improve conciseness: state the backup protocol once (Safety Rules) instead of restating it in Phase 2's Auto-Backup section, explain the S1-S4 model once and reference it thereafter, and consolidate the per-phase reference-loading lists into the single Lazy Loading policy.
Fix the internal inconsistency between the 'AI-Flavor Detection (16-Item Checklist)' heading in Phase 3 (and 'Re-run the 16-item AI-flavor checklist' in Step 5.1) and the correct 24-item checklist (1-16 English, 17-24 Chinese) used in Phase 4 and defined in references/common/ai_flavor_checklist.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most content is skill-specific operational knowledge (NSFC golden-ratio page budgets, the four-paragraph closure model, ROI scoring table, agency review criteria) that Claude could not reconstruct, but there is real redundancy: backup rules appear in both "Safety Rules" and "Auto-Backup & Checkpoints", the S1-S4 planning model is fully explained twice (Phase 1 Step 1.3 and the Phase 2 planning protocol), lazy-loading instructions are repeated per phase, and the full GRANT_STATE.json schema is inlined. This fits anchor 3 ("mostly efficient but includes some unnecessary explanation or could be tightened") better than anchor 4, where over-explanation would be only minor. | 3 / 5 |
Actionability | Guidance is highly concrete for an instruction-only skill: executable script invocations ("python3 scripts/validate_length.py <proposal_dir> --mode cn|us --json"), sentence-level templates ("The long-term goal of [PI] is... The objective of this application is..."), quantified thresholds (30-50 references, 400-character abstract, 5% golden-ratio deviation), a full state schema, and a P0/P1/P2 severity taxonomy. It falls short of anchor 5 because Phase 2's core instruction "Load the appropriate template from templates/" cannot be executed — the templates/ directory and config.yaml do not exist in the bundle, so a meaningful fraction of the concrete guidance dead-ends (though a documented fallback to built-in knowledge partially mitigates this). | 4 / 5 |
Workflow Clarity | The six phases each have explicit Entry Criteria, numbered workflow steps, and Exit Criteria, plus a command table for phase jumping, an error-handling section, and partial-use rules. Validation is layered exactly as anchor 5 requires: deterministic script checks with a manual fallback feedback loop ("If scripts are not available or fail, perform these checks manually"), Tier 2 semantic checks, simulated review, and a final compliance checklist; destructive operations are guarded by auto-backup, copy-before-write, and confirmation rules. | 5 / 5 |
Progressive Disclosure | Structure is strong: a dedicated "Reference Files" section enumerates the bundle, references are one level deep and clearly signaled per phase, and an explicit lazy-loading policy ("Do NOT read all reference files at once... reduces token usage by ~60%") is provided; all references/ and scripts/ paths verified to exist. It stays at anchor 4 rather than 5 because the body references files that are absent from the bundle — templates/us/*, templates/cn/* (7 files), config.yaml, and CREDITS.md — which breaks navigation for the drafting and final-formatting phases, a gap beyond the 'minor organization gaps' that anchor 5 tolerates. | 4 / 5 |
Total | 16 / 20 Passed |