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.
The body is an exemplary operational protocol: phased routing, executable prompts and commands, hard validation gates with rollback, and explicit refusal handling for every shortcut a user might request. Its weaknesses are token efficiency — rules are restated across Constants, Key Rules, Refusals, and phase text — and progressive disclosure, since everything lives inline in one long file while all referenced documents sit outside the bundle where their existence cannot be confirmed.
Suggestions
Consolidate the rule statements that currently appear in Constants, Key Rules, Refusals, and the phase bodies (e.g., THRESHOLD immutability and blind-sub-agent enforcement each stated three times) into a single authoritative 'Key Rules' section referenced from the phases.
Move the worked examples with real data — the Phase 3 sample ranking table, Phase 8 console report, and B3 boundary report — into a reference file (e.g., examples.md) and keep only the output format skeleton inline, cutting a substantial fraction of the ~470 lines.
Verify or inline the externally-referenced material (bump-validation-protocol.md, observation-lifecycle.md, rubric-memo.template.md, cheat-score-blind SKILL.md): these paths resolve outside the skill bundle, so a consumer cannot confirm they exist; either bring the needed protocol steps into the bundle's references/ or restructure so the skill is self-contained.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is ~470 lines where the same rules appear three times: THRESHOLD=0.8 is stated in Constants ("这条写死"), Phase 3 ("不允许临时调低"), and Key Rules #2; blind-sub-agent enforcement appears in Phase 2, Refusals, and Key Rules; Phase B rules are restated in "Phase B 不做的事". Not 2: nothing explains concepts Claude already knows — all material is domain-specific protocol, not padding; not 4: the duplication between Constants / Key Rules / Refusals and the long sample console outputs (Phase 8, B3) could be consolidated without losing meaning. | 3 / 5 |
Actionability | Quotes: the verbatim sub-agent spawn prompt ("script_path: <prediction header 的 Script Path>…不要询问用户 —— 你没有用户"), the exact leak-guard command "grep -E '\d+\s*[wWmMkK万]|播放|实绩|实际'", exact bucket formulas ("buckets = { '退步': (-inf, baseline * 0.3), … }"), concrete state-file JSON, and the full external-LLM audit prompt — all copy-paste executable. Not 4: even edge handling (refusals, retries "重发 Task 最多 3 次", abort conditions) is specified with concrete actions. | 5 / 5 |
Workflow Clarity | Quotes: a routing diagram (Phase A0) dispatching to Phase B vs Phases 0-8, hard validation gates ("排序一致性 < THRESHOLD（默认 0.8） → 本地拒绝"), a mandatory independent audit (Phase 4) with conflict handling ("本地 PASS + 外部 REJECT → 视为 REJECT"), a post-write self-check with rollback (Phase 5d leak guard → "abort bump + 回滚"), and a failure-mode table with retry/exclusion logic. Despite being a destructive batch operation (rewrites rubric_notes.md and all prediction files), every step has an explicit checkpoint or abort path, which clears the destructive-batch cap. | 5 / 5 |
Progressive Disclosure | All links point outside the bundle ("[shared-references/bump-validation-protocol.md](../../shared-references/bump-validation-protocol.md)", "[cheat-score-blind](../cheat-score-blind/SKILL.md)", "[templates/rubric-memo.template.md](../../templates/rubric-memo.template.md)") and no references/, scripts/, or assets/ directories exist in the skill — so the referenced material is unverifiable from the bundle. Meanwhile content that belongs in separate files is inlined: full protocol detail, sample re-scoring tables with real data, complete console-report examples, and three bucket-derivation formula blocks. Not 2: headers, a flow diagram, and tables make it navigable, not a header-less wall; not 4: the 470-line monolith plus external-only references leave real organization gaps. | 3 / 5 |
Total | 16 / 20 Passed |