Content
63%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-organized, knowledge-dense reference with genuinely useful curated parameters, model recommendations, and decision tables, plus a clear three-stage pipeline and staged-metric validation guidance. Main weaknesses: textbook explanations Claude already knows inflate token cost, no explicit feedback loop for tuning, and every referenced external path is absent from the skill bundle.
Suggestions
Remove or compress known-concept explanations (the king-queen vector analogy, why RAG exists, the inverted-index-as-book-index analogy, the headhunter/interviewer metaphor) — keep only the non-obvious curated details like parameter values and failure-mode judgments.
Verify or fix the companion paths: chapter3/retrieval-pipeline/, chapter3/dense-embedding/, chapter3/sparse-embedding/, and book/chapter3.md do not exist in the skill bundle — either ship them in references/ or remove the dangling pointers.
Add an explicit tuning feedback loop, e.g., 'run the test-query set, report recall@k per stage; if dense recall is low → adjust chunk size/overlap, if fusion is weak → check RRF k, if top-k precision is low → add/strengthen the reranker' to turn the implicit validation guidance into a validate→fix→re-measure workflow.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient and contains high-value curated specifics ("每块 256-1024 token、重叠 10%-20%", "常用 k1=1.5、b=0.75", "DF > N/2 时取值为负，实现需设下限"), but it also spends tokens on concepts Claude already knows: the king-queen analogy ("'国王' - '男性' + '女性' ≈ '女王'"), why RAG exists ("模型训练数据有截止日期"), what an inverted index is ("如同书末术语索引页"), and the headhunter/interviewer metaphor. This matches 'mostly efficient but includes some unnecessary explanation that could be tightened' rather than the noticeably-padded level 2 or the minor-trim level 4. | 3 / 5 |
Actionability | Guidance is concrete and executable in intent: parameter starting points ("重叠 10%-20%", "k1=1.5、b=0.75", "RRF…k 常取 60", "候选池前约 50 个用跨编码器"), named models (BGE-M3, bge-reranker-v2-m3), a query-type decision table, and companion code directories with described functionality. As an instruction-only skill the absence of inline code is not penalized, but there are minor gaps — no inline commands and the referenced code paths are not part of the skill bundle — placing it just below fully copy-paste-ready guidance. | 4 / 5 |
Workflow Clarity | Multi-step processes are clearly sequenced: "分块 → 索引 → 检索 → 注入生成" and "并行检索 → 结果融合 → 神经重排序", with a staged-validation checkpoint ("流水线每一级（稠密、稀疏、融合、重排）都单独报告指标，才能定位是哪一级丢的召回"). However, the fix-and-retry feedback loop is only implicit (hints like "检索质量不达标时定位调优（分块大小、融合方式、是否加重排序）" exist but no explicit validate→adjust→re-measure loop), matching 'clear sequence with most checkpoints; minor validation gaps'. | 4 / 5 |
Progressive Disclosure | Structure is good: well-organized sections (何时使用/核心原则/实践模式/常见陷阱/配套代码/深度阅读) with references clearly signaled one level deep in a dedicated section ("chapter3/retrieval-pipeline/ — 稠密 + 稀疏 + RRF 融合…完整流水线"). However, none of the referenced paths (chapter3/*, book/chapter3.md) exist in the skill bundle — there are no references/, scripts/, or assets/ directories — so the pointers are dangling within the skill, and some dense inline parameter detail could live in reference files. This sits between the well-structured anchor 4 and anchor 3's problematic references; the clearly signaled but unverifiable references keep it at the lower bound of 4 rather than 5. | 4 / 5 |
Total | 15 / 20 Passed |