ai21.net
语言

任务分派主题动态短片

演示视频

任务分派主题动态短片 — 1

提示词

You're directing a 20-second animated short about the Claude Haiku 5.5 release. You're the director, the art department, the animator, the sound designer and the QA team, and you have real tools to call.

I'm giving you the idea, the tools and the quality bar. Everything else, from the beats to the look to the jokes, is yours to design. I trust your taste more than a beat sheet I could write. Your capabilities are far beyond what you assume, so work like it: plan carefully, push the craft, and use the tools and the time this deserves. Spend generously where it improves the film and economically everywhere else.

Final output: `out/level3.mp4`, 1920×1080, 30fps, exactly 20.0 seconds, with voices, music and sound design. At the end, you'll also save your process as a reusable skill.

---

## 1. The idea

**Opus is overloaded with work it shouldn't be doing.** It's buried in small, repetitive tasks: summaries, classifications, lookups, compactions. Then it has a realization: it can hand all of that to Haiku. It calls Haiku, Haiku arrives fast and clears the pile, and Opus is finally free to do the hard, important work it's built for.

That's the premise. How you tell it is up to you: the staging, gags, rhythm, camera, character beats, what the eureka moment looks like, how Haiku arrives, and how the delegation plays out. Sonnet can appear if it earns its place.

**Fixed points:**
- 20.0 seconds, 16:9.
- Opus and Haiku are characters in the same family as the Claude character.
- The film must read with the sound off.
- The ending respects reality: Haiku takes the volume, and Opus keeps the hard thinking. Anthropic says Opus 5.5 and Sonnet 5.5 remain the better choice for complex agentic coding, so Haiku is never shown replacing Opus.

---

## 2. Style reference

The reference is Jake Moran's "Animated agent session story": https://www.prompt-motion.com/jake11moran-a269c4. It was made with this skill: https://github.com/heygen-com/hyperframes-community-skills/tree/master/skills/session-story. Read the skill, including its references on motion and the room.

Take from it:
- The hand-drawn watercolour-on-paper look.
- A cozy room as a world.
- Cartoon blocks with faces as characters.
- Physical paper props that carry the story (paper planes as tasks).
- A score that's cued to every action.

Don't copy it. Use it as a starting point and push it further: richer staging, better acting from the characters, more cinematic camera, and a stronger hook. It should look like it belongs to the same family, made by someone who studied it closely.

---

## 3. Facts (lock these before any creative work)

Fetch https://www.anthropic.com/claude-haiku-5-5 and write `facts.json`: every claim you might use, each with the exact supporting sentence from the page. Only claims in `facts.json` may appear on screen or in dialogue.

This is a story first, so use one or two facts at most, woven into the world rather than presented as slides. For example, a price tag being re-stamped, or Haiku's speed shown rather than stated. The film must end on a card with **Claude Haiku 5.5**.

---

## 4. Audience and hook

The audience is AI and developer Twitter/X: people who use Claude every day and know the model names and the jokes about them. Before writing the treatment:
- **Research:** spend a little time on how people currently talk about Opus, Sonnet and Haiku, and what's being said about this launch. Find one or two references or running jokes this audience would recognize and enjoy. Use them only if they fit the story naturally. Never use real people or brand logos.
- **The hook:** the first second has to work as a still frame on a busy timeline. Someone scrolling should instantly see "a character drowning in work" and want to see what happens.
- **Text on screen:** this is motion design that tells a story, not a lyric video. Use text sparingly and only where it adds something: a beat the picture can't carry alone, the fact stamp, and the end card. Integrate it into the world (painted, stamped, written in marker) rather than overlaying it.

---

## 5. Tools

The keys are in `.env`. Pick the best tool for each job, and tell me which you chose and why.

| Tool | Key | Use it for |
|---|---|---|
| Web search and fetch | — | Source page, style reference, audience research |
| Higgsfield (connector) | — | Image generation; call `models_explore` first to pick the best model for flat illustrated characters. Also music or sound effects if its audio models are the best fit. |
| GPT Image (OpenAI) | `OPENAI_API_KEY` | A second image model. Test both on the same prompt and keep whichever matches the style better. |
| ElevenLabs | `ELEVENLABS_API_KEY` | Character voices, sound effects, and music (Eleven Music) |
| Remotion | — | The film itself |
| ffmpeg + Python | — | Audio analysis, frame extraction, checks |
| Subagents | — | Parallel research and independent review |

---

## 6. Order of work

### Step 1: Treatment
Write `treatment.md`:
- Your take on the story in one paragraph.
- The beat sheet with timings.
- What happens in each shot: camera, acting, the gag.
- Where the music hits.
- What each character says.
- Which fact appears and how.
- Why this version is the strongest way to tell it.

Write two or three distinct treatments, then pick one and explain why.

### Step 2: Style bible
Write `style/bible.md`:
- Palette: 6 colours maximum, with hex values and a reason for each.
- Ink line rules: weight, and the "boiling" redraw jitter.
- Watercolour and paper texture.
- Lighting.
- Type: local font files only.
- Character design rules for Opus and Haiku: size ratio, shapes, how each one moves, and their expression set.

### Step 3: Key art (images as the base, no video generation)
The generated images are your key frames and design references. You are the in-betweener. Generate:
1. **Style frames:** 3 options of the world. Pick one.
2. **Character sheets:** Opus and Haiku, each with front, three-quarter and side views and their full expression set. Use the chosen style frame as a reference so everything shares one world.
3. **Key poses:** one image per key pose in your treatment (roughly 6–10), each built from the character sheets and the style frame, matching the camera and composition of the shot.

Put everything on one contact sheet.

**Stop and wait for my approval** before building anything.

The images are references. The final film is drawn in code to match them, and generated images never appear on screen. Trace the shapes, poses, proportions and composition from the key art into SVG, and animate between them.

### Step 4: Audio first
The film's timing comes from the audio.
1. **Voices:** audition at least 2 ElevenLabs voices per character. Opus should feel warm, low and deliberate. Haiku should feel quick, bright and eager. Generate each line separately, with 2 takes, and keep the best delivery.
2. **Music:** generate a track that fits your treatment, about 22s long, with its structure matching your beats: for example, a drop when Haiku arrives. If the generated track won't land on your beats, trim and offset it with ffmpeg rather than changing the picture.
3. **Sound effects:** a sound for every physical action. Paper, whooshes, impacts, the lightbulb, stamps.
4. **Timing map:** detect the music's beat onsets and write `timing.json`. It holds every beat time, the drop frame, each voice line's start and end frames, and every key action snapped to the nearest beat. All animation timing reads from `timing.json`, with nothing hardcoded.
5. Play me the audio-only mix before building the picture.

### Step 5: Build in Remotion
- **Components:** one per character, prop and set piece. Story beats live as data in `story.json`.
- **Drawing:** watercolour fills via SVG turbulence and displacement filters, a multiply-blended paper texture, and boiling ink lines.
- **Motion craft:** anticipation before every big action, follow-through after, squash and stretch, springs for landings, arcs for every moving object, and no linear easing.
- **Acting:** the characters have to perform. Give them readable poses, holds where the audience needs to read a face, and timing that's funny.
- **Mouths:** mouth shapes driven by the voice track's loudness. They must be still when the character is silent.
- **Camera:** push-ins, a whip or a shake where it helps. Motivate every camera move.
- **Mix:** music, voices and sound effects on separate layers. Duck the music under dialogue.

---

## 7. Verification loops (do not skip any)

1. **Key-pose match:** at every key pose, render the frame and compare it side by side with its key art. Write what differs and fix it.
2. **Sound-off read:** render a silent cut. A fresh subagent that has seen nothing else reviews a contact sheet at 2 frames per second and retells the story. If it can't, the staging failed. Fix it and repeat.
3. **Sync check:** list every catch, landing, stamp and dialogue line with its actual frame vs its `timing.json` target. Everything must be within 1 frame.
4. **Fact check:** every word on screen and in dialogue, matched to `facts.json`. Cut anything unmatched.
5. **Technical:** duration exactly 20.0s, stereo audio, peaks below −1 dBFS, no unintended dead air.
6. **Director's review:**
- Watch the whole film at least twice as frames.
- Write a brutal critique as the most demanding motion designer you know: what feels assembled, what's slow, what doesn't land.
- Fix the top 3 issues, then review again.
- Only then hand it over, with a 10-frame contact sheet.

---

## 8. Package it as a skill

Save the full pipeline as `.claude/skills/launch-film/SKILL.md`, so that next time *"Make a launch film for [URL] about [idea]"* is enough.

The skill should include:
- The fact lock
- The treatment process
- The style bible as a reference file
- The key-art process with both image models
- Voice auditions
- Music and sound effects
- The timing map
- The reusable Remotion components
- Every verification loop

---

## Approval gates
1. `facts.json` and the chosen treatment
2. Style bible, character sheets and key poses → **wait for my approval**
3. Audio-only mix and `timing.json`
4. Final film, contact sheet and critique
5. Skill

Keep a task list throughout. At each gate, show me the work, not a description of it.

中文译文

你正在执导一部 20 秒的动画短片,主题是 Claude Haiku 5.5 的发布。你同时是导演、美术部门、动画师、音效设计师和 QA 团队,而且你有真实可用的工具可以调用。

我给你想法、工具和质量标准。其余的一切,从节拍到视觉风格再到笑点,都由你来设计。比起我自己能写出的节拍表,我更相信你的品味。你的能力远远超出你自己的假设,所以就按这个水准去工作:仔细规划,把工艺推到极致,用好工具,也用好这件事值得投入的时间。凡是能提升影片的地方就大方投入,其余地方则精打细算。

最终输出:`out/level3.mp4`,1920×1080,30fps,精确 20.0 秒,包含配音、音乐和音效设计。最后,你还要把你的流程保存成一个可复用的 skill。

---

## 1. 想法

**Opus 被大量本不该它做的工作压垮了。** 它埋没在琐碎重复的任务里:摘要、分类、查询、压缩。然后它意识到:它可以把这些统统交给 Haiku。它调用 Haiku,Haiku 迅速赶到,清空了积压,Opus 终于腾出手来做它生来就该做的、困难而重要的工作。

这就是前提。怎么讲由你决定:场面调度、搞笑桥段、节奏、镜头、角色节拍、灵光一现的时刻长什么样、Haiku 怎么登场,以及这次任务委派如何展开。如果 Sonnet 配得上出场,它可以出现。

**固定要点:**
- 20.0 秒,16:9。
- Opus 和 Haiku 是与 Claude 角色同属一个家族的角色。
- 影片在静音状态下也必须看得懂。
- 结尾要尊重现实:Haiku 承担数量,Opus 保留硬核思考。Anthropic 说,在复杂的 agentic coding 上,Opus 5.5 和 Sonnet 5.5 仍然是更好的选择,所以绝不能让 Haiku 取代 Opus。

---

## 2. 风格参考

参考作品是 Jake Moran 的 “Animated agent session story”:https://www.prompt-motion.com/jake11moran-a269c4。它是用这个 skill 做的:https://github.com/heygen-com/hyperframes-community-skills/tree/master/skills/session-story。读一读这个 skill,包括里面关于动效和房间的参考。

从中借鉴:
- 手绘水彩纸感。
- 以一个温馨的房间作为世界。
- 带面孔的卡通积木作为角色。
- 推动叙事的实体纸道具(纸飞机代表任务)。
- 每个动作都有对应的配乐提示。

不要照抄。把它当作起点,再往前推:更丰富的调度、更好的角色表演、更有电影感的镜头,以及更强的钩子。它应该看起来同出一门,而且是一个仔细研究过参考的人做的。

---

## 3. 事实(在任何创作之前先锁定)

抓取 https://www.anthropic.com/claude-haiku-5-5,写出 `facts.json`:你可能用到的每一条说法,都要附上页面上精确的支撑句。只有 `facts.json` 里的说法才能出现在画面或对白里。

这首先是一个故事,所以最多用一两条事实,把它们织进世界里,而不是像放幻灯片一样摆出来。比如,一个被重新盖章的价签,或者把 Haiku 的速度“演”出来,而不是“说”出来。影片必须以一张写有 **Claude Haiku 5.5** 的卡片收尾。

---

## 4. 受众与钩子

受众是 AI 和开发者 Twitter/X:每天都用 Claude、熟悉模型名字以及关于它们的梗的人。在写阐述稿之前:
- **调研:** 花一点时间看看现在大家怎么谈论 Opus、Sonnet 和 Haiku,以及关于这次发布都有哪些说法。找一两个这个受众能认出来并会心一笑的典故或常驻梗。只有当它们能自然融入故事时才用。绝不要使用真实人物或品牌 logo。
- **钩子:** 第一秒必须在繁忙的信息流里作为一张静止画面就成立。刷到的人应该一眼看出“一个被工作淹没的角色”,并想知道接下来会发生什么。
- **画面上的文字:** 这是用动态设计讲故事,不是歌词视频。文字要克制,只在能加分的地方使用:画面单独承载不了的一个节拍、事实印章,以及片尾卡。把文字融进这个世界(画上去、盖上去、用马克笔写上去),而不是直接叠在画面上。

---

## 5. 工具

密钥都在 `.env` 里。为每项工作挑选最合适的工具,并告诉我你选了什么、为什么。

| 工具 | 密钥 | 用途 |
|---|---|---|
| 网页搜索与抓取 | — | 来源页面、风格参考、受众调研 |
| Higgsfield(连接器) | — | 图像生成;先调用 `models_explore`,为扁平插画角色挑选最合适的模型。如果它的音频模型最合适,也可用于音乐或音效。 |
| GPT Image(OpenAI) | `OPENAI_API_KEY` | 第二个图像模型。用同一个提示词测试两者,保留更贴合风格的那个。 |
| ElevenLabs | `ELEVENLABS_API_KEY` | 角色配音、音效和音乐(Eleven Music) |
| Remotion | — | 影片本身 |
| ffmpeg + Python | — | 音频分析、抽帧、检查 |
| Subagents | — | 并行调研与独立审阅 |

---

## 6. 工作顺序

### 第 1 步:阐述稿
写出 `treatment.md`:
- 用一段话讲你对这个故事的处理。
- 带时间点的节拍表。
- 每个镜头里发生什么:镜头、表演、笑点。
- 音乐在哪些点落下。
- 每个角色说什么。
- 哪条事实出现,以及怎么出现。
- 为什么这个版本是讲这个故事最有力量的方式。

写两到三个不同的阐述稿,然后选一个,并解释原因。

### 第 2 步:风格圣经
写出 `style/bible.md`:
- 色板:最多 6 种颜色,给出十六进制值和每种颜色的理由。
- 墨线规则:粗细,以及“沸腾”式的重绘抖动。
- 水彩与纸张纹理。
- 灯光。
- 字体:仅使用本地字体文件。
- Opus 和 Haiku 的角色设计规则:体型比例、形状、各自的运动方式,以及表情组。

### 第 3 步:主视觉(以图像为基础,不用视频生成)
生成的图像是你的关键帧和设计参考。你来当中间画师。生成:
1. **风格帧:** 3 个世界的方案。选一个。
2. **角色设定图:** Opus 和 Haiku,各含正面、四分之三侧面和侧面视图,以及完整表情组。用选定的风格帧作为参考,让所有内容共享同一个世界。
3. **关键姿势:** 阐述稿里的每个关键姿势一张图(大约 6–10 张),每张都基于角色设定图和风格帧构建,并与该镜头的机位和构图匹配。

把所有内容汇总到一张总览图上。

**停下来,等我批准**,然后再开始搭建任何东西。

这些图像只是参考。最终影片是用代码绘制出来与它们匹配的,生成的图像绝不会出现在屏幕上。把主视觉里的形状、姿势、比例和构图描摹成 SVG,然后在它们之间做动画。

### 第 4 步:音频优先
影片的时间节奏来自音频。
1. **配音:** 每个角色至少试听 2 个 ElevenLabs 声音。Opus 应该温暖、低沉、从容。Haiku 应该轻快、明亮、跃跃欲试。每句台词单独生成,每条录 2 遍,保留最好的一版演绎。
2. **音乐:** 生成一段符合你阐述稿的曲子,约 22 秒长,结构要与你的节拍对应:比如 Haiku 登场时来一个 drop。如果生成的曲子对不上你的节拍,就用 ffmpeg 裁剪和偏移它,而不是改画面。
3. **音效:** 每个物理动作都要有声音。纸张、呼啸、撞击、灯泡、盖章。
4. **时间映射:** 检测音乐的节拍起点,写出 `timing.json`。它记录每个节拍时间、drop 所在的帧、每句配音的起始与结束帧,以及每个吸附到最近节拍的关键动作。所有动画的时间都从 `timing.json` 读取,不得硬编码。
5. 在搭建画面之前,先放给我听纯音频混音。

### 第 5 步:在 Remotion 中搭建
- **组件:** 每个角色、道具和场景单元各一个。故事节拍以数据形式存放在 `story.json` 里。
- **绘制:** 通过 SVG 湍流和置换滤镜实现水彩填充,正片叠底混合纸张纹理,以及沸腾的墨线。
- **运动工艺:** 每个大动作前有预备,动作后有跟随,挤压与拉伸,落地用弹簧,每个移动物体都走弧线,不使用线性缓动。
- **表演:** 角色必须真的在演。给他们清晰可读的姿势,在观众需要读懂表情的地方留住,以及好笑的节奏。
- **嘴型:** 嘴型由配音轨道的响度驱动。角色沉默时嘴必须不动。
- **镜头:** 推镜、甩镜或震动,在有用的时候用。每一个运镜都要有动机。
- **混音:** 音乐、配音和音效分轨。对白时把音乐压低。

---

## 7. 验证闭环(一个都不能跳过)

1. **关键姿势匹配:** 在每个关键姿势处渲染该帧,与对应的主视觉并排比较。写下差异并修正。
2. **静音可读性:** 渲染一个无声版本。让一个全新的、没看过任何其他内容的 subagent 以每秒 2 帧查看总览图,并复述故事。如果它复述不出来,就是调度失败。修正后重来。
3. **同步检查:** 列出每个接住、落地、盖章和对白句的实际帧与 `timing.json` 目标帧。所有误差必须在 1 帧以内。
4. **事实核查:** 屏幕上和对白里的每个字词都要与 `facts.json` 对应。对不上的删掉。
5. **技术:** 时长恰好 20.0 秒,立体声,峰值低于 −1 dBFS,没有意外的空白。
6. **导演审阅:**
- 以逐帧方式把整部影片至少看两遍。
- 以你认识的最苛刻的动态设计师的身份写一份毫不留情的批评:哪里像是拼凑出来的,哪里拖沓,哪里没打中。
- 修正最重要的 3 个问题,然后再次审阅。
- 到这时才交付,并附上一张 10 帧总览图。

---

## 8. 把它打包成一个 skill

把完整流程保存为 `.claude/skills/launch-film/SKILL.md`,这样下次只要说 *“为 [URL] 做一部关于 [idea] 的发布短片”* 就够了。

这个 skill 应包含:
- 事实锁定
- 阐述稿流程
- 作为参考文件的风格圣经
- 使用两个图像模型的主视觉流程
- 配音试音
- 音乐与音效
- 时间映射
- 可复用的 Remotion 组件
- 每一个验证闭环

---

## 审批关卡
1. `facts.json` 和选定的阐述稿
2. 风格圣经、角色设定图和关键姿势 → **等我批准**
3. 纯音频混音和 `timing.json`
4. 最终影片、总览图和批评意见
5. Skill

全程维护一份任务清单。每到一道关卡,给我看作品本身,而不是描述。

溯源

原作者
kelanoo
发布日期
2026-10-08(由推文 ID 反解)
原始出处
查看原帖 ↗
收录来源
zhuyansen/awesome-opus-5.5-video · 2108199183983497433 · 未声明许可