Content
85%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.
This is an exceptionally actionable, well-sequenced tool reference: concrete commands, explicit validation gates, and honest failure semantics throughout. Its main weaknesses are inlined per-command detail that should live in reference files for a bundle this size, and some prose/table redundancy that inflates token cost without adding guidance.
Suggestions
Move the exhaustive pip-compose and deck-compose flag documentation into references/pip-compose.md and references/deck-compose.md, keeping only a usage example and the key invariants in SKILL.md.
Trim the final subcommand table rows that merely restate the prose sections (e.g. assemble, pip-compose, deck-compose) to a one-line pointer, or conversely compress the prose and let the table carry the detail.
Consider per-domain reference files (e.g. references/audio.md, references/compose.md) for the stage-by-stage workflows, keeping SKILL.md as the stage map plus common invocations.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, procedural, and assumes competence — it never explains what ffmpeg, LUFS, or SRT are, and nearly every sentence carries a concrete constraint or path. The deduction is for redundancy: pip-compose and assemble each get long prose treatment and then reappear summarized in the final subcommand table, and some run-on sentences (e.g. the assemble --video-only paragraph) could be tightened without losing information. | 4 / 5 |
Actionability | Guidance is fully executable: copy-paste-ready bash invocations for pip-compose, normalize, burn-srt, duck, denoise, and interp; exact flag semantics with defaults and valid ranges (--size 默认 .22，范围 .1–.5); exit-code contracts (0/1/2); and a complete per-subcommand table of inputs and output artifacts. This matches the 'copy-paste ready commands covering the common cases' anchor. | 5 / 5 |
Workflow Clarity | The multi-stage pipeline (Stage 1–14) is explicitly sequenced, with two gates (GATE A after Stage 5, GATE B after Stage 9), checkpoint semantics ('产物文件存在性即 checkpoint'), dedicated validation commands (visual-plan-check, deck-boundaries, edit-audit with 'status=pass 才交付'), and explicit error-recovery loops ('常规输入错误先自纠' with retry guidance and 旁路条件 that fail loudly rather than degrade silently). | 5 / 5 |
Progressive Disclosure | Section structure is clear and the single reference (references/edit-audit.md) exists and is clearly linked, but nearly all per-command detail is inlined in a ~22 KB SKILL.md: the 12-line pip-compose flag reference, the deck-compose mode matrix, and the full 30-row subcommand table would all sit more appropriately in per-command reference files for a bundle shipping 40 scripts. This fits 'some structure but could be better organized; content that should be separate is inline' rather than the good-structure anchor. | 3 / 5 |
Total | 17 / 20 Passed |