Agent Skills: Seedance 2.0 Prompt Guide

Seedance 2.0 Prompt Guide

Overview

Create Seedance 2.0 prompts using the official BytePlus Dreamina guidance. Optimize for natural-language logic, explicit reference mapping, audiovisual synchronization, and predictable edits.

For exact templates, limits, and task-specific phrasing, read references/prompt-guide.md.

Workflow

1. Classify the request:
   • New generation: T2V, I2V, R2V, or V2V.
   • Text rendering: slogan, subtitle, or speech bubble.
   • Reference transfer: image subject, multi-image elements, video motion, camera motion, or VFX.
   • Video editing: add, remove, modify, extend forward/backward, or complete tracks.
2. Map all provided materials before writing:
   • Use explicit identifiers: Image 1, Image 2, Video 1, Audio 1.
   • If sequence matters, tell the user or assume assets are uploaded in the intended order.
   • State what each reference controls: subject identity, composition, outfit, logo, motion, camera movement, effects, audio, or scene layout.
3. Start the prompt with the logical core:
   • Required: subject plus action, i.e. who does what.
   • Add environment and aesthetics only after the core action is clear.
   • Add camera movement, cuts, timing, and audio when they affect the output.
4. Make reference instructions explicit:
   • Prefer "Use the composition of Image 1", "Maintain the character design from Image 2", "Match the motion of Video 1", or "Keep the VFX trajectory from Video 1".
   • For multiple references, bind each reference to one job when possible.
5. Make text and speech renderable:
   • Use common words and short phrases.
   • Quote exact on-screen text or dialogue.
   • Specify timing, position, entrance/appearance behavior, and visual style only when needed.
   • For subtitles, require bottom-center placement and synchronization with audio rhythm/pacing.
6. Preserve continuity for edits:
   • For removals, say to keep the rest of the video unchanged.
   • For replacements, say to preserve original motion and camera work.
   • For extensions, describe content before/after the input clip without regenerating the original segment.
7. Return the result in the form the user needs:
   • If the user asks for a prompt, provide one polished prompt.
   • If the user asks for variants, provide distinct prompt options with short labels.
   • If reviewing, list issues first, then provide an improved prompt.

Prompt Shape

Use this structure unless the user's format requires otherwise:

[Mode/task]: [subject + action]. [Reference bindings]. [Environment and aesthetic]. [Camera movement/cuts/timing]. [Text, dialogue, subtitle, or speech-bubble instructions]. [Audio or sound design]. [Continuity/edit constraints].

Keep prompts concise but concrete. Seedance 2.0 follows natural language well; do not turn the prompt into a dense parameter list unless the target API requires parameters separately.

Quality Checklist

Before finalizing a prompt, verify:

• The subject and action are unambiguous.
• Every referenced image/video/audio has an identifier and a specific purpose.
• Sequence-dependent references say "in strict order" or equivalent.
• Text rendering uses simple vocabulary and avoids unusual symbols unless required.
• Dialogue, subtitles, speech bubbles, music, and sound effects are synchronized when relevant.
• Editing prompts include timestamp/spatial location when available and preservation constraints.
• Track-completion prompts use no more than 3 input video clips and no more than 15 seconds total when the user mentions clip counts or durations.

Reference

• Source distilled from BytePlus ModelArk document 2222480, "Dreamina Seedance 2.0 series prompt guide", last updated 2026-04-23.
• Load references/prompt-guide.md for task-specific templates and official constraints.