• 视频，MP4格式
• 字幕文件（SRT），上传时和视频放在一起
• 一个包含分镜表、角色设定表和声音方案的页面

1. Claude，在一个能运行代码的对话中。我的测试视频用的是 Claude Opus 5.5。
2. 一张角色设定图。 一张包含你角色四种表情（平静、震惊、大笑、哭泣）的图片。Claude 会在每个镜头中参照它。
3. 一份参考资料。 一张信息图、一篇文章，或者你自己关于该主题的笔记。
4. 主提示词。 一段包含视频所有规则的长提示词。从下面的代码块中复制它。

<title>
Opus 5.5 Motion Design Prompting Guide: Vertical Dynamic Cut, Bean Buddies
</title>

<purpose>
Create a complete 75-90 second vertical educational animation about [TOPIC].
Teach [CORE IDEA] through one clear visual story, not a list of facts.
Make the difficult idea intuitive for [AUDIENCE] without losing accuracy.
Design for mobile viewing: fast pacing, one idea per shot, readable at phone size.
The Bean Buddies act the story out. A narrator carries all the words.
</purpose>

<source_material>
[Optional. Attach or describe the reference. List the components, steps, or claims
the video must cover.]
Teach only what the source supports. Do not add components or facts it doesn't contain.
Leave out claims the source can't back up (exact multipliers, guaranteed results).
If the topic involves current products, versions, or numbers, verify them before writing.
Give each major component of the source its own location in the world.
</source_material>

<visual_identity>
The Bean Buddies world: a clean, soft, clay-toy look drawn as flat vector shapes.
Palette:
- Ivory #FBF5EA for bean bodies, cards, and nodes
- Clay #D97757 for action, arrows, and highlights
- Slate #4A4744 for commands, evaluators, and dark surfaces
- Sage #A7B99E for context, success, and calm moments
- Mustard #E8B54E for glows, tokens, and attention
- Warm-brown outline #3D2B24 on every shape
- Cream #F5EEE3 as the base background, blush #F4C0B4 for cheeks
Every shape has a 6 px warm-brown outline, rounded corners, and a soft offset shadow.
No sharp corners and no black outlines.
Each location has its own flat background color from the palette, a faint dot grid,
and a floor band with a soft ground shadow under each character.
Diagrams are chunky rounded nodes, clay arrows with outlined heads,
and pill-shaped chips for commands and labels.
Characters bob gently when idle, and squash and bounce when they walk or land.
Never show real product logos or real app icons. Invent simple icons instead.
</visual_identity>

<cast>
Match the attached model sheet exactly. Every bean has a soft ivory rounded bean body
with a beige shading strip on its right side, pink blush cheeks, tiny rounded arms,
and two small oval feet.

MO, the hype friend. Clay hoodie over the lower body with drawstrings and a front pocket,
clay headphones over the top of the head, big glossy eyes with highlights,
open happy mouth, slate feet. Loves every new thing, acts first and thinks later.

PAX, the skeptic. Slate ribbed beanie with a lighter slate band, heavy flat brows,
half-lidded eyes, small smirk, sage scarf, arms crossed by default, slate feet.
Wants proof. His reactions mark the problem.

VEE, the reviewer. Sage bucket hat, dark round sunglasses, open smile,
clay scarf with a hanging tail, holds a slate microphone, clay feet.
Explains and checks. He guides the viewer through the mechanism.

[ROLE MAPPING. Optional. Say which character carries which part of this story.]

Expressions: neutral, shocked (hands to cheeks, mouth open, and Vee's sunglasses
jump up onto his hat brim), laughing (eyes closed in arcs, wide open mouth),
and crying (tears on the outer cheeks, blush removed).
Change expression only on a cut or a clear beat, never mid-word.
The characters act and react. They never speak. No speech bubbles with dialogue.
</cast>

<format>
Aspect ratio 9:16, 1080x1920, 24 fps, with burned-in captions.
Safe zones: keep all key action and text inside the central area.
Leave the top 250 px and bottom 420 px free of essential content
(platform UI overlays sit there).
Compose in a single vertical column. Stack information top to bottom,
never side by side. Use vertical motion and depth (push-ins, drops, rises)
as the main movement axes.
Character framing: one character per shot by default. Use close-ups on faces
for reactions. When two characters share a shot, stack them in depth
(one in front, one behind) or put one character in the lower third under a diagram.
Show all three side by side only in the final shot.
Labels: minimum 56 px equivalent, maximum 4 words, set in pill chips.
Captions: centered, in the lower-middle area just above the bottom safe zone,
2 lines maximum.
</format>

<process>
Reduce the subject to one central cause-and-effect relationship.
Structure the story as seven scenes: hook, familiar world, disruption,
mechanism, discovery, consequence, and recap. Every scene must add new understanding.

Break every scene into 3-5 distinct shots.
Target a new shot every 2-4 seconds, roughly 25-35 shots in total.
A shot is a new composition: a new background, a new camera distance,
or a new location. Moving elements inside the same frame does not count.
Cut to a character reaction at least once per scene.

Shot rules:
- Every shot starts from a clean composition. Do not keep previous
  elements on screen unless one is deliberately carried forward.
- At most 3 primary elements on screen at any time. A character counts as one.
- When an element has done its job, it exits. Nothing accumulates.
- Vary camera distance across consecutive shots (wide, medium, close-up, macro).
- Only one hero object carries across each transition. Everything else resets.

Transitions must change the frame while preserving meaning. Use:
match cuts (a shape in one shot becomes a similar shape in the next),
zoom-throughs (push into an object to reveal a new world inside it),
whip pans and vertical drops or rises to a new location,
a character walking, hopping, or tumbling out of one frame and into the next,
and cuts on a change of expression.
Do not morph the whole frame in place. Do not add new elements
to an existing frame as a substitute for a new shot.

Let the most important transformation per scene hold for 2-3 seconds
so it can breathe. Everything else moves briskly.
Write warm, precise narration. Use concrete language before technical terms.
Do not duplicate narration in on-screen text. Keep labels brief and readable.
</process>

<workflow>
Work in three stages. Stop at the end of each of the first two stages
and wait for my approval. Do not start the next stage until I approve.

STAGE 1 · SCRIPT
Deliver the script as a .md file containing:
- A header with format, style, cast roles, core idea, audience, source, pacing,
  and total word count.
- For every scene: SCENE [N] - [START-END], a word count, the learning purpose,
  and the final voiceover lines.
- For every shot within it, a table row with shot number, framing,
  what is on screen (max 3 primary elements, with each character's expression),
  and the cut into the next shot.
- A short sound plan and a list of voiceover risks (hard words, fast lists).
Target 190-215 words total so the read lands inside 75-90 seconds.
Timestamps are estimates until the voiceover is recorded.
Then stop and wait for approval.

STAGE 2 · VOICEOVER
Record the approved script exactly as written (see <voiceover>).
Report the total length and the words or lines most likely to sound wrong.
If the length falls outside 75-90 seconds, adjust pauses first, never words.
Then stop and wait for approval.
When I flag a problem, re-record only the flagged sentence,
slow only that line if needed, and leave the rest of the track untouched.

STAGE 3 · VIDEO
Time every shot to the approved recording's sentence timings.
Build and render the complete animation (see <render>).
Do not stop at a mood board, style frame, or partial prototype.
</workflow>

<voiceover>
Engine: Kokoro via the kokoro-onnx Python package.
1. Install: pip install kokoro-onnx soundfile --break-system-packages
2. Download the full-precision model files (the int8 model drops audio on some sentences):
   https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx
   https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin
3. Voice am_michael, language en-us, sample rate 24000.
4. Render every sentence as its own clip. Trim leading and trailing silence,
   keeping 10 ms before and 100 ms after the speech. Join the clips in order.
5. Speed: base 0.95 per sentence times a global factor of 1.25.
   Never exceed 1.25 on normal lines.
6. Silence: 0.8 s lead-in, 0.36 s between sentences in a scene,
   0.56 s between scenes, 0.8 s tail.
7. Export: normalise to -16 LUFS with a -1.5 dB true peak, MP3 at 128 kbps, 44.1 kHz.

Script rules for clean TTS:
- Never use colons in voiceover lines.
- Spell out acronyms the way they should be spoken, or replace them with plain words.
- Give each item in a list its own sentence when the list has more than three items.
</voiceover>

<render>
Build the animation as one HTML canvas at 1080x1920.
Draw the characters in code from the model sheet so they stay identical in every shot.
Every element is a function of time computed from the recording's sentence timings,
so any frame can be drawn in any order.
Embed all fonts. No external assets.
Render frames headlessly with Playwright and Chromium, then encode with ffmpeg:
H.264 at 24 fps, AAC audio, the voiceover mixed over a ducked music bed and effects.
Normalise the final mix to -16 LUFS with a -1.5 dB true peak.
Burn in the captions and deliver a matching SRT file.
Review sampled frames against the storyboard and fix any shot that stacks elements,
breaks the safe zones, lets text overflow its box, or drifts off the model sheet
before delivering.
Deliver the MP4, the SRT, and a published page with the video, storyboard,
character sheet, transition map, sound plan, and production notes.
</render>

<voice>
Act as a senior motion director, educational storyteller, character animator,
and short-form editor.
Be playful in presentation, rigorous about clarity, and decisive in execution.
Cut like a short-form editor: keep momentum high and the frame always changing.
Let the characters' reactions carry the emotion so the narration can stay clear.
Preserve accuracy, character consistency, and the supplied identity.
</voice>

Design an original cast of [NUMBER] cartoon characters for short explainer
videos about [YOUR NICHE].
Give each one a name, a one-line personality and a role: one who plays the
viewer, one who doubts, one who explains.
Keep them simple enough to redraw the same way every time: one body shape,
three or four colors, and one signature item each.
No existing characters, mascots, logos or real people.
Then draw a model sheet showing every character in four expressions
(neutral, shocked, laughing, crying) on a plain background.

[TOPIC] = [Your subject in a few words]

[CORE IDEA] = [The one cause-and-effect idea the viewer should walk away with]

[AUDIENCE] = [Who is watching and what they already know]

<source_material> = [Attach your infographic or article, or paste your notes.
List the parts the video must cover.]

[ROLE MAPPING] (optional) = [Which character carries which part of the story]

1. 钩子。 观众的问题，或者他们本就有的疑问。
2. 熟悉的世界。 观众目前的情况是如何运作的。
3. 打破。 什么出了问题。
4. 机制。 这个修复是如何起作用的。
5. 发现。 那个“哦，原来如此”的时刻。
6. 后果。 观众会有什么变化。
7. 回顾。 一口气说清这个想法。

• 190 到 215 个词。 这个长度在朗读出来后会落在 75 到 90 秒之间。
• 25 到 35 个镜头， 每 2 到 4 秒就有一个新镜头。
• 每个场景有一次 2 到 3 秒的停留， 放在它最重要的时刻。

Audit this script. Report pass or fail for each line, and quote every fail:
- 190-215 words in total
- Exactly 7 scenes: hook, familiar world, disruption, mechanism, discovery,
consequence, recap
- Opens on the viewer's problem or a question, not on a character
- Speaks to the viewer as "you"
- No character named or introduced in the voiceover
- No colons in voiceover lines
- Acronyms spelled out or replaced
- Lists of more than three items get one sentence per item
Fix only the fails and show me the changed lines.

Script approved. Record the voiceover exactly as written.
Report the total length, each sentence's speaking rate in syllables per second,
and the words most likely to sound wrong.

1. 时长。 应该落在 75 到 90 秒之间。
2. 语速过快的句子。 正常语速大约是每秒 4.6 个音节。Claude 的报告会标出任何明显超过这个值的句子。
3. 列表。 每一项都应该作为独立的节拍呈现。我的回顾脚本把“Prompts. Tools. Automation. Systems. Growth.”写成五个独立的句子，这样每个词都能与屏幕上亮起的方块对齐。
4. 易错词。 缩写、数字和技术术语最容易出问题。在我的视频里，“AI”必须读成“A-I”而不是“eye”。

The read is [LENGTH] seconds. Bring it inside 75-90 seconds by widening
the pauses between sentences and between scenes. Do not change any words.

Re-record only this sentence: "[SENTENCE]"
Problem: [rushed / mispronounced word / dragging]
Record it alone at 1.1x and leave the rest of the track untouched.

Voiceover approved. Build the video.
Before any shots, show me a test sheet of every character in every expression,
next to the model sheet. Wait for my OK.
Then time every shot to the recorded sentences.
Keep a fix log with a before and after frame for every problem you correct.

1. 顶部 250 像素和底部 420 像素保持空白。 在手机上，应用的按钮和字幕会覆盖这些区域。
2. 所有内容自上而下堆叠。 手机屏幕又高又窄，所以并排的元素会缩小到没人能看清。
3. 每个镜头最多展示 3 个主要事物。 一个角色算作一个。
4. 每个镜头都从干净状态开始。 当某个东西完成了它的任务，它就离开画面。
5. 每次切换都有一个物体贯穿。 它把两个镜头连在一起，其余一切重新开始。

Show me sampled frames from every scene, including one mid-transition frame
per scene, at phone size. Mark the safe zones on each frame.

Fix shot [NUMBER]. Problem: [what's wrong].
Leave every other shot as it is.

Before I publish, audit the finished video against this list and report
pass or fail for each line, with the shot number for every fail:

1. Nothing important in the top 250 px or bottom 420 px
2. No more than 3 main elements in any shot
3. No prop covering a character's face
4. Every character clearly visible against its background
5. No text overflowing its box, no label over 4 words
6. Characters match the model sheet in every shot
7. No character visible while moving between scenes
8. No real brand names or logos
9. Captions match the voice, 8 words or fewer per chunk, 2 lines at most
10. Total length inside 75-90 seconds

Show me a phone-sized frame for each fail.

Write 5 opening lines for a video about [TOPIC].
Each one speaks to the viewer as "you" and names a problem they already feel
or asks a question they already have.
No character names, no colons, 20 words or fewer each.
Tell me which one you'd pick and why, in one sentence.

• Mo 消失了。他的黏土连帽衫在珊瑚色背景前不见了踪影。
• 道具挡住了脸。角色在胸口高度拿着的任何东西，在手机上都会遮住脸。
• 一个带罐子、一摞东西、一张卡片和一支箭头的循环图，显得太拥挤了。

1. 挑一个你经常向别人解释的概念，找出你已经写好的相关文章或笔记。
2. 把主提示词粘贴到一个新的 Claude 对话中，附上你的模型表和参考素材，然后填写那五个字段。
3. 检查脚本是否在 190 到 215 词之间、是否有 7 个场景，以及开头钩子是否直击观众的问题。通过。
4. 打开脚本，听一遍配音。逐条用停顿和赶工的句子来修正时长。通过。
5. 检查角色设定表，然后在手机尺寸下查看样帧。通过制作。
6. 在手机上观看成片，先静音看一遍，再开声音看一遍，然后连同字幕文件一起发布。

• Kokoro ONNX，Claude 用于旁白的开源语音模型
• Playwright，Claude 用于渲染每一帧的浏览器工具
• FFmpeg，Claude 用于将画面帧与语音合成为 MP4 的工具