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.
The body delivers dense, high-value operational knowledge with concrete rules, contrast tables, and a verification checklist — clearly the work of someone with real production experience. Its weaknesses are re-explanations of transformer fundamentals Claude already knows and a broken progressive-disclosure story: the referenced companion files are absent from the bundle and substantial detail (vendor policies, research frontier) is inlined rather than split out.
Suggestions
Cut or compress the explanations of concepts Claude already knows — the KV Cache mechanism intro, the quadratic-prefill/linear-decode cost aside, and the causal-attention justification — down to one framing sentence each, keeping only the operational implications.
Fix progressive disclosure: either ship the referenced files (chapter2/kv-cache/ experiment code, book/chapter2.md) in the bundle or remove/repair the dangling '配套代码' and '深度阅读' references so every path resolves.
Move the per-vendor CoT-return policy table and the research-frontier paragraph (Models Take Notes at Prefill, RoPE re-location) into a one-level-deep reference file (e.g. references/vendor-policies.md), leaving a short summary and a clearly signaled pointer in SKILL.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Several passages re-explain transformer fundamentals Claude already knows: 'KV Cache 把前文 token 的键值对（K/V）缓存下来', '无缓存时 prefill 注意力计算量随上下文长度平方级增长', '因果注意力决定每个 token 的 KV 只依赖它之前的 token', and the long research-frontier paragraph. Most other content (vendor-specific CoT policies, quantified failure cases) is genuinely non-obvious, so this matches anchor 3 ('mostly efficient but includes some unnecessary explanation or could be tightened') rather than 4. | 3 / 5 |
Actionability | Guidance is concrete and checkable: three explicit iron rules, the layered layout diagram, vendor-specific API contracts ('必须原样回传 reasoning_content，否则 API 直接返回 400'), and a copy-ready checklist ('[ ] system prompt 与 tools 顺序在代码中是常量'). Matches anchor 4 (mostly executable guidance with minor gaps, e.g. no command for reading cached_tokens); as an instruction-only skill with actionable guidance it is not penalized for lacking code. | 4 / 5 |
Workflow Clarity | The flow 何时使用 → 核心原则 → 实践模式 → 常见陷阱 → 落地检查清单 forms a clear diagnostic/application sequence, and the checklist provides explicit verification checkpoints ('会话恢复时能重建与缓存字节流一致的消息序列'). Matches anchor 4 (clear sequence, most checkpoints present); not 5 because there is no error-recovery feedback loop, and not 3 because checkpoints are explicit rather than implicit. | 4 / 5 |
Progressive Disclosure | Section headers are well organized, but the referenced paths 'chapter2/kv-cache/', 'book/chapter2.md' and 'context-compression' do not exist in the bundle (no references/, scripts/, or assets/ directories are present), so the references are dangling. Additionally the ~90-line body inlines content (per-vendor CoT return policies, research-frontier notes) that belongs in one-level-deep reference files. Matches anchor 3 ('some structure; references present but not clearly signaled; content that should be separate is inline'), not 4 because the dangling references are not usable navigation. | 3 / 5 |
Total | 14 / 20 Passed |