短视频口播文案检查
Overview
This skill is a diagnostic tool, not a rewriting tool. It takes a creator's self-written voiceover draft and returns a structured health-check report: which parts work, which parts fail, and why—graded item by item (✓ / △ / ✗) with the exact evidence sentence and a clear verdict.
The division of labor is deliberate: the AI points out the problems; the creator decides what to fix. Diagnosing is reliable and value-certain; rewriting is creative and high-variance, and a dissatisfying rewrite would only damage trust in the tool. So this skill stops at the report. No rewritten script, no "polished version", no five-part deliverable.
Use this skill whenever a user asks to review, health-check, or diagnose a short-video script against a voiceover checklist; to learn how to write hooks, the five-stage structure, or the persona point-of-view rules; or to find out "where is this draft weak".
When to use
- "体检这篇口播稿" / "体检这篇脚本" / "这篇稿子哪里有问题" / "帮我看看这篇文案"
- "开头钩子是不是不行" / "钩子怎么写才留人" / "五段式怎么排"
- "人称是不是讲教味太重" / "结尾太像口号" / "这条视频文案诊断"
Pipeline (fixed order)
- 读稿 (Read) — sentence by sentence, mark: the opening promise, which sentence carries the strongest material, filler sentences, internal jargon, typos, and the nature of the ending.
- 骨架五查 (Skeleton 5 checks) — promise delivered? strongest line position? listed-but-not-explained? jargon/typos? ending = slogan or action? (See
references/self-draft-and-rules.md.) - 三关 (Three gates) — 钩子关 (hook) → 结构关 (five-stage structure) → 文案关 (density / coherence / rhythm). (See
references/hooks-and-structure.md.) - 事实核对 (Fact check) — flag every number and claim that needs verifying before publishing.
- 输出体检报告 (Report only) — present the judgment using the fixed two-part format in
references/copy-healthcheck-19.md(19-item ✓/△/✗ list + 体检信息). No rewritten script is produced.
Core rules (always enforce)
- 只体检，不改写 — 报告里绝不出现"改法"、改写句、"第一句改成…"这类替用户动笔的内容。只给 ✓/△/✗ ＋ 证据句 ＋ 判决；修改由创作者自己完成。用户若想了解某条怎么改，可在对话中另行询问，或自行查阅 references/ 对应文件。
- Persona point-of-view (人称尺) — "我" : "你" ≥ 2 : 1; "你" must never land on an imperative / promise / judgment sentence; "我" must be followed by something the creator actually did. (We flag violations; we don't rewrite.)
- One scene per day / one turn per script — no turn = no publish; a feeling must hang on a real event.
- Numbers must be true; no preset outcomes — flag any number that isn't verified and any claimed result that hasn't happened; one fact gets one number per video.
- Light call-to-action only — a soft follow-prompt is allowed; hard begging and gated bait are forbidden; at most 2 CTAs; avoid hard-selling "free" on screen.
Resources
references/hooks-and-structure.md— 八型钩子 (8 hook types), 钩子自检三问, 四条钩子铁律, 五段式 (five-stage structure), 中段小钩, 三关制总览, 新号期数据判读.references/copy-healthcheck-19.md— 文案体检表 19 项 (the 19-item copy checklist) with grading (✓ / △ / ✗), the verdict line, and the fixed report format (展示格式).references/self-draft-and-rules.md— 骨架五查 (+第6/7条), 人称尺, 一天一幕, 内容配方三条腿, 对标纪律, 截图证据, 幽默四招.
How to apply
When the user submits a draft or material:
- Read + 骨架五查 + 三关 + 事实核对, then present the judgment in the chat — a checklist with ✓／△／✗ per item, the evidence sentence, and a verdict.
- Report only. Use the fixed two-part format from
references/copy-healthcheck-19.md: (1) the 19-item list, one line per item; (2) 体检信息 — 稿名 · 字数/秒数 · 总判 · 非✓项证据 · 小账 · 判决 · 最疼的一处. - Never produce a rewritten script, a polished draft, a change-comparison table, a five-part deliverable, or any "改法"/reworded-line suggestion inside the report. The report ends at the verdict. If the user later asks for fix direction, you may point them to the relevant reference file's rule or discuss it—but never output a rewritten script.
- Load a reference file only when needed; keep this SKILL.md lean. references 是给创作者自学的写作/修改教材——报告输出时只引用其中的判断标准，不得把"怎么改"直接写成改写句抛给用户。
