How to Use the Seedance 2.0 Capability Map for Prompt Design: A Technical Guide

The Seedance 2.0 capability map is a four-layer verification framework in references/capability-map.md that separates what the model can do from what any single surface exposes, enabling precise prompt design that exploits official strengths while working around documented limits.

The Capability Map is the central reference document for the Emily2040/seedance-2.0 repository. It defines what the Seedance 2.0 video generation model can do, how to access those capabilities through specific syntax, and what limitations must shape your prompt architecture. Unlike provider documentation that makes universal claims, the map forces you to verify capabilities across four independent layers before committing to a design strategy.

The Four Verification Layers

According to the source code in references/capability-map.md lines 11-14, every capability must pass through four verification stages:

Layer	Question Answered	Evidence Required
Model capability	Can the model do this at all?	Model-level documentation describing the modality
Surface access	Does this specific surface expose it?	Availability per operation, not guaranteed across siblings
Request syntax	How is the capability expressed?	Example showing the authoring form
Returned adherence	Did the output actually follow it?	Observed take as evidence, not guarantee

This structure prevents the common failure mode of assuming a feature works everywhere just because it appears in official documentation.

Designing Into Model Strengths: Extraction Moves

The capability map enumerates concrete extraction moves you should preload before drafting any prompt. These are located in references/capability-map.md starting at line 30.

Multi-Shot Generation

Shot 1:/2:/3: labels · one action · one camera · standard tier [field] · 10-15 s / auto · shots × seconds budget

This extraction move from line 34 enables multiple shots in a single API call—a core efficiency feature of Seedance 2.0.

# Prompt

Shot 1: A close-up of a ceramic teapot steaming, the camera dollies forward, soft piano notes rise.
Shot 2: The teapot tilts, pouring steaming tea into a glass, the camera tracks left, a faint chime echoes.
Shot 3: A hand lifts the glass, sunlight glints on the surface, the camera pans up to reveal a sunrise.
duration: auto

Why this works:

• Shot 1:/2:/3: labels satisfy the Request syntax layer
• One primary action + one camera move per shot follows the Model capability extraction move
• duration: auto lets the model allocate roughly 4-6 seconds per shot, respecting the budget guideline

The exact syntax for these labels is defined in references/multishot-grammar.md line 5, which the capability map references as the authoritative request format.

Native Synced Audio

Name specific sounds, keep dialogue short, prioritize SFX > music > dialogue

From line 35 of the capability map, this move ensures audio elements render correctly:

# Prompt

Shot 1: A city street at dusk, a single car horn blares, the camera pans right to follow the car.  
Audio: "car horn" – crisp, short, on-screen.  
Shot 2: Neon signs flicker, a soft synth pad rises, the camera tilts up.  
Audio: "synth pad" – sustained, low-key.
duration: auto

Explicit Audio: tags implement the extraction move. The capability map emphasizes keeping dialogue minimal and ranking sound types by reliability.

First/Last Frame Locking

Lock endpoints, prompt → travel → resolve; use transformations & match-cuts

From line 39, this technique constrains generation between fixed visual states:

# Prompt

Start frame: a handheld shot of a wooden door, light spilling through the crack.  
Shot 1: The door opens slowly, camera tracks forward, a creak is heard.  
Shot 2: Inside, a candle flickers, the camera tilts down, soft wind whistles.  
End frame: the candle blows out, darkness settles, the camera pulls back.
duration: 12s

The First/last frame extraction move requires explicit Start frame: and End frame: declarations. Each intermediate shot maintains single-action/single-camera discipline per line 40's Literal camera verbs rule.

Physics-Based Description

Use physical verbs & consequences, not pose adjectives

Line 41 of the capability map notes that describing what happens (verbs + physical consequences) outperforms static pose descriptions for motion coherence.

Style-Specific Modes

Mode	Extraction Move	Source
2D/anime	Cel-style grammar, no lens/DOF talk	Line 44
Multilingual	Anchor texture/mood in Chinese, keep reference tags exact	Line 46

Designing Around Model Limits

When chaining generations or pushing duration boundaries, the capability map at line 50 recommends specific mitigations:

• Limit clip length to avoid continuity drift across generated segments
• Preserve exact reference roles (e.g., "motion only, no appearance")
• Re-anchor on schedule at the scene's chain-depth cap
• Treat surface-specific duration caps as facts of that surface, not of the underlying model

These constraints reflect the Returned adherence layer—what you observe in actual takes, not what documentation promises.

Integration With the Design Pipeline

The capability map sits at the foundation of Seedance 2.0's documented workflow. SKILL.md line 50 shows the "Capability check" step explicitly loading the map before any shot planning occurs. README.md line 69 positions it as the first of three design aids alongside the fidelity-allocation model and model-mechanics reference.

The complete pipeline runs:

1. Consult the Capability Map → identify extraction moves
2. Allocate fidelity (references/allocation-model.md) → distribute rendering budget
3. Apply concrete grammar (references/multishot-grammar.md) → implement syntax
4. Generate → execute API call
5. Validate adherence → use prompt-lint and eval-run tools

Summary

• The capability map (references/capability-map.md) provides a four-layer verification framework that separates model capabilities from surface-specific availability
• Extraction moves like multi-shot labels, audio tags, and endpoint locking translate capabilities into concrete prompt syntax
• The request syntax layer depends on references/multishot-grammar.md for exact formatting rules
• Design-around strategies for chaining and duration limits are based on observed behavior, not documentation claims
• The map is explicitly loaded during the "Capability check" phase documented in SKILL.md

Frequently Asked Questions

What is the difference between "official" and "[field]" markers in the capability map?

Official capabilities are documented by the model provider and generally available across surfaces. [field] markers indicate capabilities verified through production use but potentially subject to surface-specific restrictions. According to references/capability-map.md, you should verify [field] capabilities through the Surface access layer before depending on them.

Can I use multi-shot prompts on any Seedance 2.0 API surface?

Not necessarily. The capability map's Surface access layer requires explicit verification. While the Model capability layer confirms multi-shot generation exists, individual surfaces may not expose it. Check references/api-status.md for surface-specific resolution and feature availability before designing your prompt architecture.

Why does the capability map distinguish "one take is evidence, not a guarantee"?

This reflects the Returned adherence verification layer. A single successful generation proves a capability can work under specific conditions, but stochastic variation means subsequent runs may differ. For production reliability, you should accumulate multiple observed takes or implement the design-around strategies documented at line 50 of the capability map.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}