# Seedance 2 — Reference Hub

Folder: `references/ai_visual/seedance_2/`
Status: living reference, v1 (2026-05-21)
Source of truth: Higgsfield official prompting guides (Apr 2026) + ByteDance ModelArk docs. See `vocabulary.md` for full citations.

## What's here

| File | Audience | Purpose |
|---|---|---|
| `index.html` | Nikita + team | Visual dictionary — what Seedance 2 actually understands. Open in browser. |
| `examples/case_western_saloon.html` | Team | Worked case study: deconstruction of a real prompt, side-by-side with canonical rewrite. |
| `schema.md` | Other Claude chats | Canonical prompt structure as machine-readable spec. Linked from project chats. |
| `vocabulary.md` | Other Claude chats | Full lexicon table: which words Seedance honors / partially honors / ignores. |
| `README.md` | Both | This file. |

## How to use this from another chat

When you're in a project chat and need to draft a Seedance 2 prompt, paste this at the top of the request:

> Use the Seedance 2 reference base at `/Users/nikitareplianskii/Dropbox/Claude/The Core/references/ai_visual/seedance_2/`. Read `schema.md` first for canonical structure, then `vocabulary.md` for working terms. Follow the structure exactly. Do not invent field labels.

That gives Claude in the project chat enough to write a prompt that matches Higgsfield's expectations, without re-deriving the rules from memory.

## Canonical structure — TL;DR

```
HEADER       Total: <duration> / <shot_count> / <aspect_ratio>
STYLE        Aesthetic + grade + lens-feel + grain. Negative constraints.
SCENE        One prose paragraph: who, where, when. Starting mise-en-scène.
CAST         @Char1 = role, @Char2 = role, …  +  @Environment if applicable
Shot 1:      Action + camera in plain language. Dialogue inline with [emotion]:.
Shot 2:      …
…
SFX:         List of concrete sound events.
```

Full details in `schema.md`. Visual examples in `index.html`.

## Hard rules (the cheapest mistakes to avoid)

1. **Do not invent field labels** like `Static Description:` / `Dynamic Description:` / `Starting Composition:` — these are not in Seedance's grammar. Use `Shot 1:` / `Shot 2:` / `…`.
2. **Do not pack more shots than the duration allows.** Roughly 1 shot per 2–3 seconds. 5 shots = 15s, not 5s.
3. **Do not write camera body / lens model names** (Arriflex, Cooke, etc.). Use look buckets (`ARRI ALEXA aesthetic`, `85mm telephoto feel`, `anamorphic film grain`).
4. **Do not pair head movement with dialogue in the same Shot.** Speaking beat = static character. Movement beat = no dialogue.
5. **Do not repeat the same camera spec** four times. One mention is enough.

## When this file goes stale

Update when:
- Higgsfield ships a new Seedance version with breaking changes to prompt structure.
- A new tested case study lands in `examples/`.
- Vocabulary observations drift (a previously-ignored term starts working, or vice versa).

Track changes in this README at the top (status + date).