Turn "what you want to say" into "a script you can read to camera, precise to the second." Two layers of work:First de-AI-ify (so it can be read), then define the shots (precise).
Layer one: De-AI-ify (can it be read aloud)
A voiceover script and an article are not the same thing. An article can be read; a voiceover has to bespoken. AI-written scripts sound "fake" when read aloud, with consistent symptoms:
| AI-sounding symptom | Fix |
|---|---|
| Long sentences with nested clauses, can't finish in one breath | Break into short sentences, one idea per sentence |
| "Not only... but also..." "First, second, finally" written connectives | Switch to spoken transitions: "then," "you see," "here's the problem" |
| Triple parallelism, neat antithesis | Break the neatness, leave some looseness and colloquialisms |
| Em dashes, parenthetical asides | Change any punctuation that can't be read aloud into spoken insertions |
| Stock phrases like "it is worth noting," "in summary," "this means" | Delete outright or replace with plain speech |
| Piling up adjectives instead of being specific | Replace with specific numbers, examples, actions |
There is only one criterion: read it aloud. Anywhere you stumble, run out of breath, or sound like you're reading a script is a place that needs fixing.
If the user has their own "AI-sounding blacklist / review standards," read it in first and go through it accordingly.
Layer two: Shots (precise to the second)
Don't allow stopping at a vague description like "I want to make a 3-minute product video." Keep pressing with follow-up questions until every second is planned.
Follow-up question checklist (vague → precise)
- Target duration? Platform? Aspect ratio (vertical 3:4 / horizontal 16:9)?
- What's the hook in the first 3 seconds?
- How many segments? What's the one core sentence for each?
- Visual source for each segment: talking head / screen recording / B-roll animation / screenshot?
- Visual theme: color scheme, fonts, motion style?
Output:script.md(shot by shot, precise to the second)
# 视频脚本 · <标题>
平台: 视频号竖屏 3:4 | 目标时长: 90s | 风格: 深色/极简
| 时间 | 口播（去AI味后的台词） | 画面 | 呈现方式 | 转场 |
|------|----------------------|------|---------|------|
| 0:00-0:03 | 「我这半天，几乎没碰键盘。」 | 真人口播 | 特写+气泡出镜 | 硬切 |
| 0:03-0:10 | … | 录屏 | 屏幕放大高亮 | 叠化 |
| 0:10-0:18 | … | B-roll | motion 文字动画 | 推入 |
Each line must clearly state:what these seconds say, what the visuals are, how it's presented, how it cuts to the next shot. Subtitle text, text animation points, and B-roll needs are all noted here for downstream space-video-broll / space-video-subtitle to use directly.
Handoff
- Upstream:
space-video-topic's正文.mdor user draft - Downstream:
script.mdB-roll needs marked →space-video-broll; subtitle text marked →space-video-subtitle; voiceover lines → recording/editing
Orchestration (reuse existing capabilities)
- Deep de-AI-ifying rewrite → go through the user's
Humanizer-zh - Voiceover-style conversion, long-form to voiceover → go through the user's
podcast-script-generator
This skill is responsible for combining "de-AI-ify" and "precise shots" into a directly shootable script.md。
Core principles
- First make it readable, then make it precise: If you can't get rid of the AI flavor, even accurate storyboards are useless
- Force it down to the second: Don't accept vague descriptions; keep probing until every second has a plan
- Leave traces of a real person: Moderate colloquialisms, pauses, and self-corrections are more convincing than neatness