liangdabiao/make-prompt-seedance2

What it is

This repository is a practical guide for writing structured prompts for Seedance 2.0 — ByteDance's AI video-generation model. The core deliverable is a single guide file (structured-prompt.md) containing 16+ prompt templates and 8+ worked examples, plus a folder of third-party tutorial articles (docs/) and a CLAUDE.md for setting up Claude Code (an AI coding assistant) to use this project directly.

Think of it as a recipe book: instead of guessing how to describe a video to the AI, you pick a template, fill in the blanks, and get a consistent, high-quality prompt.

Who it's for

The README lists six audiences: short-video creators, self-media bloggers (product promos, talking-head videos, music-synced clips), designers/product managers, e-commerce and home-decor folks (product showcases, virtual home tours), educators (teaching videos, science animations), and film/animation people.

The tool it targets: Seedance 2.0

Three traits are highlighted:

• Has 'director thinking' — it automatically plans shots and scene arrangement.
• Has built-in world knowledge — no need to describe every physical detail (e.g., how noodles are pulled); just state the intent.
• Multi-modal references — you can attach up to 9 images, 3 videos, and 3 audio files as references simultaneously.

Core philosophy: write intent, not details

The central idea is "写意图、不写细节" (write intent, don't write details). When the AI has world knowledge, describe the outcome you want, not the step-by-step process.

• ✅ Recommended: "Generate a refined, high-end Lanzhou beef-noodle ad; pay attention to shot arrangement."
• ❌ Not recommended: describing every kneading, pulling, and pouring step in detail.

The prompt template

A structured prompt has four labeled blocks:

【风格】(style, duration, aspect ratio, mood)
【时间轴】(timeline: "0-X秒：镜头 + 画面 + 动作 + 特效")
【声音】(music + sound effects + dialogue)
【参考】(references: @图片1 ____, @视频1 ____)

The guide also includes a shot-language quick-reference table, an atmosphere-keyword library, and multi-modal reference syntax (e.g., @图片1 作为首帧 — "image 1 as first frame", @视频1 参考运镜方式 — "video 1 as camera-movement reference").

How to actually use it (5 steps)

1. Learn the model — read the overview in the repo.
2. Pick a platform — three official access points are listed: 即梦 (Jimeng, web), 小云雀 (XiaoYunque, app), 豆包 (Doubao, app; partial roll-out).
3. Choose a template — the README maps common goals to template numbers (brand ad → template 2, product showcase → template 8, short-drama clip → template 7, fight scenes → templates 3/10, home-tour → template 9, talking-head → template 11, music beat-sync → template 12, video extension → template 5).
4. Write the prompt — fill in the four-block format.
5. Upload and generate — attach reference materials, reference them with @ names, set duration and aspect ratio, then generate.

Shortcuts and workarounds

• One-sentence generation: if the subject is already in the AI's world knowledge (e.g., a known restaurant brand, a fitness exercise, a historical period), and you don't need character consistency or a special camera style, a single sentence suffices.
• Copyright avoidance: writing a well-known character's name may get the request rejected; instead use generic placeholders like "Figure 1" / "Figure 2" and describe their appearance.

Hard limits (important)

Item	Limit
Real-person face material	❌ Not supported
Total reference files	≤ 12 (≤9 images + ≤3 videos + ≤3 audio)
Video duration	≤ 15 seconds
Audio format	MP3 only
Output resolution	Mostly 720P currently

Guide structure

structured-prompt.md has numbered chapters: core philosophy → system role definition → detail-mining prompts → the 16+ templates → shot-language reference → atmosphere keywords → multi-modal syntax → special cases → 8+ full examples → output format → advanced tips/FAQ.

Practical notes

• The README links to a real demo video (on Bilibili) showing actual results, and to a separate sister project for video-style cloning.
• Version history: v1.0 (base), v2.1 (added 6 templates, 3 advanced examples, tips and FAQ).
• License: "for learning and exchange only" — not a commercial license.

Bottom line: if you want to generate AI videos with Seedance 2.0 but find prompt-writing hit-or-miss, this repo hands you a structured, copy-paste-able system with templates, an example bank, and explicit constraints — so you can get consistent output without reinventing the prompt format.