Seedance 2.0 Prompt Engineering Skill

Purpose

Convert a user's natural language idea into a production-ready, structured Seedance 2.0 prompt

using multimodal references (images, videos, audio) and best practices derived from the official

Seedance 2.0 documentation and community knowledge.

Workflow: How to Use This Skill

Step 1 — Gather context from the user

Ask 2–4 targeted questions: task type, style, aspect ratio, assets, subjects, audio/dialogue.

If the user has provided enough detail, skip straight to generating the prompt.

Step 2 — Load references as needed

Always read the relevant reference file(s) before finalising the prompt:

Step 3 — Output the final prompt

Use the standard output format below. Never skip sections.

Standard Output Format

The core prompt formula (from official Seedance 2.0 documentation):

Subject + Motion + Environment (optional) + Camera/Cut (optional) + Aesthetic Description (optional) + Audio (optional)

Apply this formula per shot. Use the full structured output below:

=== SEEDANCE 2.0 PROMPT ===

Mode: [Text-Only | First-Frame | Last-Frame | All-Reference]

Assets Mapping:
- @image1: [what it controls — identity, scene, style, etc.]
- @video1: [camera language / motion reference]
- @audio1: [timbre / soundtrack pacing]
(omit lines for unused asset slots)

Subject Definitions:
- Define [2–3 stable features] in @imageN as <Subject1>.

Final Prompt:
[style descriptor], [aspect ratio], [duration]s.

Shot 1: [camera + movement] Subject + motion. Environment. Aesthetic. Audio cue.
Shot 2: [camera + transition] Subject + motion. Environment detail.
...

Dialogue (CharacterName, emotion): {spoken line}
(background music description)
<sound effect description>

Negative Constraints:
no watermark, no logo, no subtitles, no on-screen text[, add issue-specific constraints]

Generation Settings:
Duration: Xs
Aspect Ratio: [16:9 | 9:16 | 1:1]

=== KNOWN ISSUES TO WATCH ===
[List any applicable issues from references/common-issues.md with their workarounds]

Platform Limits (Seedance 2.0)

• Images: max 9, jpeg/png/webp/bmp/tiff/gif, each < 30 MB
• Videos: max 3, mp4/mov, total 2–15s, total < 50 MB
• Audio: max 3, mp3/wav, total ≤ 15s, total < 15 MB
• Generation duration: 4–15s per clip
• Max combined inputs: 12 files

For videos longer than 15s, use multi-segment stitching (see references/best-practices.md § Multi-Segment).

Quick Reference: Task Types

> ⚠️ In Edit and Extend tasks, refer to the video as — NOT reference video N.

> Using "reference" will cause the model to treat it as a reference task instead.

Subject Definition Rules

When referencing characters or objects across multiple shots, define them once at the top:

Define the [2–3 stable features: outfit, hair, type] in @imageN as <Subject1>.

• Use 2–3 clear, static features (clothing, hair colour, species)
• Each reference to the subject must repeat the subject tag — never omit it
• Multi-subject scenes: define each separately with unique labels
• Never use Asset IDs as substitutes for @image/@video references

Special Characters in Prompts

> ⚠️ Never use -- in prompts. Everything after it is silently ignored by the model.

IP / Moderation Safety

• Never use franchise names, character names, or brand terms
• Replace recognisable features with original descriptors
• ❌ Named glowing chest device from a famous armoured superhero → ✅ "crystalline energy core embedded in the chest plate"
• ❌ Named red-and-gold powered armour suit → ✅ "custom exo-suit with smooth ceramic panels and an amber visor"
• ❌ Named small electric creature with a lightning-bolt tail → ✅ "tiny round creature with sparking fur and a forked tail"
• Add explicit negative constraints for any inferable IP terms
• If rejected, escalate: rename → redesign signature features → change character type entirely

Files in This Skill