Content
67%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 well-structured orchestration skill: the review pipeline is sequenced with explicit failure gates, concrete script commands, and genuine one-level-deep progressive disclosure verified against the actual bundle. Its main weaknesses are boilerplate constraint sections that dilute token efficiency and a few underspecified steps (master-prompt generation, the 校验 checkpoint) plus orphaned bundle scripts.
Suggestions
Trim or externalize the generic 公共硬约束 block and de-duplicate the file listing (lines 122-126) and output-management rules, which are stated twice, to tighten conciseness.
Make the 校验 step concrete by naming the actual validation (e.g., `python3 <nsfc_reviewers_path>/scripts/validate_skill.py ...`) and specify how the master prompt is assembled from master_prompt_template.md.
Reference the orphaned bundle scripts — finalize_output.py in 输出整理 and cleanup_intermediate.py for intermediate_dir handling — so the disclosed bundle matches what the body actually uses.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The core process sections are lean and imperative, but roughly 30 lines of generic "公共硬约束" boilerplate (bug-collection policy, BAC audit, generic security rules), a duplicated bundle-file listing (lines 122-126 repeat references already cited inline), and repeated output-management rules add unnecessary tokens. Not 2 because the actual review workflow is written efficiently; not 4 because the padding exceeds minor trimmable instances. | 3 / 5 |
Actionability | Concrete guidance dominates: an executable script invocation (`python3 <nsfc_reviewers_path>/scripts/list_proposal_files.py --proposal-path <proposal_root>`), precise config keys, the `ceil(N * consensus_threshold)` consensus formula, and numbered parallel-mode steps. Not 5 because master-prompt generation is only sketched and the 校验 step names no concrete check or command. | 4 / 5 |
Workflow Clarity | A clear end-to-end sequence (前置检查 → 通读与结构化理解 → 并行/单组评审 → 聚合与排序 → 阶段判断 → 输出整理 → 校验 → 失败与恢复) with explicit failure gates (".tex 数量为 0 时直接失败") and a retry-from-failed-step recovery loop. Not 5 because the 校验 checkpoint is vague ("静态检查、脚本验证或人工复核" without naming what or how); not 3 because checkpoints are mostly explicit and the skill is read-only, so the destructive-operation cap does not apply. | 4 / 5 |
Progressive Disclosure | References are one level deep and well signaled: the body cites references/expert_*.md, references/aggregation_rules.md, references/master_prompt_template.md, scripts/list_proposal_files.py, and scripts/build_parallel_vibe_plan.py, and all of these exist in the bundle, with a dedicated file-listing section for discovery. Not 5 because three bundle scripts (finalize_output.py, cleanup_intermediate.py, validate_skill.py) are present but never referenced from the body. | 4 / 5 |
Total | 15 / 20 Passed |