# make-prompt-seedance2 and the case for describing intent to a video model

> make-prompt-seedance2 is a structured prompt guide for Seedance 2.0, ByteDance's AI video generation model, packaged as a document, a set of source articles and a reusable skill for coding agents. Its core argument is that a model with world knowledge should be given intent and shot structure rather than a detailed recipe.

**liangdabiao/make-prompt-seedance2** — 一句话：试错成本越来越高，提示词的重要性从来没有像今天这样大。seedance2结构化提示语和通用SKILL，整合了 Seedance 2.0（字节跳动 AI 视频生成模型）的结构化提示词使用方法，帮助用户快速上手并创作出高质量的视频内容。包含 16+ 个模板和 8+ 个示例 | 带货广告| TVC广告| 真人实拍视频效果

- Repository: https://github.com/liangdabiao/make-prompt-seedance2
- Website: https://seedance-video-prompt.liangdabiao.com
- Stars: 730 · Forks: 101
- Language: Unknown
- License: not declared
- Published: 2026-10-21 · Updated: 2026-10-21 · Language: en
- Canonical page: https://hysenlabs.com/projects/liangdabiao-make-prompt-seedance2

## A guide and a skill, with no code to run

make-prompt-seedance2 collects a structured method for writing prompts for Seedance 2.0, ByteDance's AI video generation model. The repository is not a library and not an application. Its primary language is recorded as unknown, it has no GitHub releases, and the licence field is empty, with the README stating that the document is for study and exchange only.

The repository layout reflects that. `structured-prompt.md` is the core guide, described as a complete tutorial on writing Seedance 2.0 prompts with more than sixteen templates and more than eight worked examples. `docs/` holds source articles collected from several named authors. `CLAUDE.md` is a working guide for Claude Code, and the `skills/` directory turns the guide into something a coding agent can use directly, with the README pointing at agents such as codex and workbuddy. There is also an `ads-prompt/` directory and a storyboard prompt document at the root.

There is no install step, because there is nothing to install. The guide is read from the `structured-prompt.md` file or from the published site, and the timeline format inside it looks like this, with each segment written as shot, image, action and effect:

```
...
```

The audience is broad and stated as a list: short video, short drama and vlog creators, self media accounts making commerce videos, talking head videos and beat-synced music clips, designers and product managers making UI promos and advertisements, e-commerce and home renovation teams making product and show flat tours, education and knowledge payment content, and film, animation and music video work.

If your interest is integrating Seedance into software, the `skills/` directory is the relevant part. If your interest is the prompting technique itself, `structured-prompt.md` is the document to read.

## Three entry points, and only one of them is fully open

The guide lists where Seedance 2.0 can be used, and the availability differs sharply between them.

The web platform at jimeng.jianying.com is fully launched and carries the full feature set. The Xiaoyunque app at xyq.jianying.com is open to all users as a download. The Doubao app is in a grey release, meaning partial availability for some users only. The guide also lists an API trial endpoint at exp.volcengine.com/ark/vision.

Those three lines determine what kind of project is realistic. A workflow that depends on batch generation through an API is a different commitment from a person typing prompts into a web page, and the guide's advice on templates and one-sentence generation reads differently depending on which of the three you can actually reach.

The README also links a separate repository, video-clone-lite, for the case where the goal is to reproduce the effect of an existing video rather than to write a prompt from scratch. The division of labour is sensible, and it is a reminder that this repository is about authoring prompts rather than about video tooling in general.

## The prompt has four parts: style, timeline, sound, references

The scaffold the guide asks you to fill in is short, and the shape of it matters more than any single template. It opens with a style line carrying the art style, the duration in seconds, the aspect ratio and the mood. It follows with a timeline where each segment is written as shot plus image plus action plus effect, with time ranges given in seconds. Then comes sound, combining the background music, the sound effects and the dialogue. The last block is references, written as an at-sign followed by a material name and a note on what it should be used for.

The timeline is the part with the most pull. Writing the video as explicit time ranges turns a single generation into a shot list, which is what makes the advice about writing intent rather than detail workable rather than a slogan. You are not describing how to knead dough; you are saying that this is a polished advertisement and asking the model to block it out.

The same intent-first principle governs the other fields. A style line of a few words carries more information than a paragraph, and a reference list of a few entries tells the model which assets exist and what role each one plays, rather than embedding a description of each asset.

## Write intent, not detail, and let the model block the shots

The guide's central concept is titled as writing intent and not writing detail, and it is justified by what the model already knows. Seedance 2.0 is described as having director-like thinking that arranges shots automatically, and as carrying world knowledge so that an encyclopedic description is unnecessary.

The guide's own contrast is between a recommended prompt, producing a polished high-end Lanzhou lamian advertisement with attention to shot arrangement, and an unrecommended one that spells out every step of kneading, pulling and ladling the soup. Both describe the same advertisement. The first is short and delegates the blocking.

A short prompt works, the guide argues, when three conditions hold: the subject has world knowledge, such as a restaurant brand, a fitness movement or a historical period; no specific character consistency is required; and no particular camera movement style is required. Give it a subject it knows, and it can direct.

The failure mode follows directly. If a character must stay consistent across shots, or the camera work must be a specific move, one sentence is not enough, and the timeline and reference blocks of the scaffold are where that requirement gets expressed. The framework is not asking for less detail everywhere, it is asking for detail in the places where detail changes the result.

## Reference syntax and the limits that will stop a plan

The reference syntax is short and positional. An at-sign plus a material name can serve as the first frame, another as the last frame, another as a character appearance reference, a video reference can supply camera movement, and an audio reference can be used for the soundtrack. The guide's point is that a reference says what a piece of material is for, not just that it exists.

The limits table is the part to read before promising a client a video. Real human face material is not supported. The total number of files is capped at twelve, made up of at most nine images, three videos and three audio files, so the listed maxima can only all be reached in combination rather than at once. Total video length is fifteen seconds or less. Audio must be MP3. Generation resolution is currently mainly 720P.

Those constraints shape what a project can be. A thirty second brand film with a presenter on camera is not a Seedance 2.0 job, and neither is a multi-shot animation with a recurring character that has to look identical every time.

## Templates by scenario, and the workaround for named characters

The guide maps templates to scenarios rather than presenting them as an undifferentiated list. Template two is for brand advertising, template eight for product display, template seven for short drama segments, templates three and ten for character fight scenes, template nine for a home renovation walkthrough, template eleven for talking head video, template twelve for beat-synced music clips, and template five for extending a video.

That mapping is the part that saves time in practice. Finding the right structure by reading sixteen templates is slow, while recognising that a task is a product display and going straight to template eight is not.

There is also a copyright workaround, and it is worth stating plainly because it is both a technical tip and a policy of the platform. Using the name of a well known character may lead to the request being rejected. The advice is not to write character names at all, to use labels such as Figure 1 and Figure 2 instead, and to describe appearance traits instead. This is the same reference discipline as the image and video references, applied to characters, and it keeps a prompt from being rejected outright.

The guide also notes that a finished example is linked in the README, so the expected output is visible before you write your own prompt.

## A ten part document, updated twice, and licensed loosely

The core document has a defined internal structure. Section zero covers the core ideas and principles of use. Section one is a system role definition with a step by step guidance flow. Section two goes into detail on content, visuals, camera work, action and sound. Section three holds the prompt formula and the templates. Section four is a camera language quick reference, section five an atmosphere keyword library, section six the multimodal reference syntax, section seven special case handling, section eight the worked examples, section nine the output format requirements, and section ten onwards the advanced material on quality traits, techniques, cautions and frequently asked questions.

That structure is a prompt system rather than a document of tips, which is what makes it usable inside an agent. The `skills/` directory packages it so a tool can apply it, and `CLAUDE.md` gives the working guidance for doing that with Claude Code.

On versions, the README records two document versions. v2.1 integrates the highlights of the source articles in `docs/`, adding six practical templates, three advanced examples, advanced techniques and a FAQ section. v1.0 was the base version. The repository's last push was on 2026-10-04, it is not archived, and it has no tagged releases, so the document version in the README is the only version marker you can rely on.

The licence statement is that the document is for study and exchange only, and the repository's licence field is empty. That is a narrower grant than an open source licence, and it is the one thing to check before copying the templates into a commercial workflow.

## Conclusion

make-prompt-seedance2 suits a creator or developer who already has access to Seedance 2.0 through one of the three platforms and wants a repeatable prompt format instead of improvised one-off descriptions, and it suits anyone wiring a coding agent to a video model through the `skills/` directory. It is the wrong tool for a team that has not settled on a platform, and it will not help with the parts the guide itself lists as hard limits, including real human faces and clips longer than fifteen seconds. Verify first which platform your account can reach, since the web version is fully available while the app version is in a partial rollout for some users, and read the limitations table in the guide before planning a project around reference counts or duration.

## FAQ

### What is make-prompt-seedance2?

It is a structured prompt guide for Seedance 2.0, ByteDance's AI video generation model, made up of a core guide in `structured-prompt.md` with more than sixteen templates and more than eight examples, source articles in `docs/`, and a `skills/` directory that packages the method for coding agents to apply directly.

### Where can I use Seedance 2.0 according to the guide?

The guide lists the web platform at jimeng.jianying.com as fully launched, the Xiaoyunque app at xyq.jianying.com as open to all users as a download, and the Doubao app as in a partial grey release for some users. An API trial endpoint is also listed at exp.volcengine.com/ark/vision.

### What are the limits of Seedance 2.0 listed in the guide?

Real human face material is not supported, the total number of files is at most twelve with up to nine images, three videos and three audio files, total video length is fifteen seconds or less, audio must be MP3, and generation resolution is currently mainly 720P.

### How should I reference images, video and audio in a Seedance 2.0 prompt?

The reference syntax uses an at-sign plus a material name, and each entry states its role, for example as the first frame, as the last frame, as a character appearance reference, as a reference for camera movement, or as the soundtrack. The same idea applies to characters, where the guide advises using labels such as Figure 1 and Figure 2 plus appearance traits instead of well known names.

### When is a one line prompt enough for Seedance 2.0?

The guide states that one line suffices when the subject is covered by world knowledge, such as a restaurant brand, a fitness movement or a historical period, when no specific character consistency is required, and when no particular camera movement style is required. Otherwise the timeline and reference blocks of the template are needed.

## Sources

- [Issues](https://github.com/liangdabiao/make-prompt-seedance2/issues)
- [liangdabiao/make-prompt-seedance2 on GitHub](https://github.com/liangdabiao/make-prompt-seedance2)
- [Project website](https://seedance-video-prompt.liangdabiao.com)
- [README](https://github.com/liangdabiao/make-prompt-seedance2/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/liangdabiao-make-prompt-seedance2