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 a dense, highly actionable operator manual: a fully sequenced 16-stage pipeline with hard gates, numeric validation thresholds, rework/escalation rules, and exact commands and paths. Its weaknesses are repeated clarifications that cost tokens and a reference layer whose target files are absent from the bundle, leaving the progressive-disclosure structure unverifiable.
Suggestions
State the "generic pipeline always applies; type workflows only cover Stage 1–2" rule once (e.g. in the three-layer section) and reference it elsewhere instead of repeating it three times, and merge the two story-develop intake explanations.
Ship the referenced files (workflows/*.md, tools/avatar/SKILL.md, references/edit-audit.md) in the skill bundle, or remove/inline the dangling references so navigation works as delivered.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is extremely dense and operational — exact paths, flags, env vars, no conceptual filler Claude already knows — but the clarification that the generic pipeline is not a fallback and that type workflows only cover Stage 1–2 is repeated three times (the three-layer table note, the Stage chain intro, and the gate discipline bullet), and the story-develop intake explanation appears twice. That is noticeable redundancy that could be tightened, matching anchor 3 rather than 4. | 3 / 5 |
Actionability | Guidance is fully concrete and executable: exact subcommand compositions ("assemble --manifest segments.json --verify-fps 25 --expect-durations slots/shotdur.json", "clip-trim --pre-buffer 0.5", "normalize --locked-audio"), exact deliverable paths, named env vars with routing and exit-2 behavior, and explicit deferral of parameter detail to tool docs. Instruction-only skill whose guidance is copy-paste actionable across the common cases. | 5 / 5 |
Workflow Clarity | The Stage 0→15 chain is explicitly sequenced with two hard gates (GATE A text, GATE B materials) that mandate stopping for approval, validation checkpoints with numeric thresholds (self-eval <3 必返工, slideshow_risk ≥4.0 fail, verdict=pass to continue), feedback loops (rework caps of 3, wall-time limits with escalation instead of silent spinning), and a decision audit trail. This matches the anchor-5 example's validate-fix-retry discipline. | 5 / 5 |
Progressive Disclosure | The three-layer structure is declared up front and references are one level deep and clearly signaled ("见 workflows/story-develop.md", "读 tools/avatar/SKILL.md", "见 references/edit-audit.md"), but none of the referenced files (workflows/*.md, tools/*/SKILL.md, references/edit-audit.md, AGENTS.md) exist in the skill bundle, so navigation cannot be verified and is likely broken as shipped. Detail that could live in reference files (Stage 12 combination recipes, env dependency routing) is also inlined, keeping this at anchor 3. | 3 / 5 |
Total | 16 / 20 Passed |