seedance-prompting-claude-skill

KO | EN

gitlite — search

#react #ai #typescript #llm #golang #deep-learning #linux #macos #java #windows #awesome #cli #kubernetes #ios

seedance-prompting-claude-skill

(★ 1)

Claude Skill for Seedance 2.5 / 2.0 prompt engineering (ByteDance, Dreamina, Jimeng, Doubao) — 6-block formula, @reference roles, staging, AI-UGC realism layer, native audio & lip-sync, debug matrix.

seedance-prompting-claude-skill Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Seedance 2.5 Prompting — a Claude Skill for ByteDance's video model

[![License: MIT](https://img.shields.io/badge/License-MIT-black.svg)](https://raw.githubusercontent.com/gbeyrouti/seedance-prompting-claude-skill/main/LICENSE)
[![Claude Skill](https://img.shields.io/badge/Claude-Skill-d97757.svg)](https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview)
[![Model](https://img.shields.io/badge/model-Seedance%202.5%20%7C%202.0-blue.svg)](https://raw.githubusercontent.com/gbeyrouti/seedance-prompting-claude-skill/main/)
[![Docs in French](https://img.shields.io/badge/docs-fran%C3%A7ais-blue.svg)](https://raw.githubusercontent.com/gbeyrouti/seedance-prompting-claude-skill/main/README.fr.md)

A [Claude Skill](https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview) that
turns Claude into a **Seedance production director**: you describe the idea, it writes a
production-grade **Seedance 2.5** prompt — right reference roles, right time staging, and the
clauses that keep the model from falling back into its default failure modes.

Works with Seedance 2.5 and Seedance 2.0 across **Dreamina, Jimeng, Doubao, BytePlus ModelArk,
Volcano Ark and fal.ai**.

> 🇫🇷 **Le skill est écrit en français** — voir le [README en français](./README.fr.md).
> Generated prompts are always in **English**, the model's prompting language.

---

## The two rules this skill encodes

Most bad Seedance output traces back to two things. Get these right and the rest is craft.

### 1. Seedance drifts toward *generic*, not toward *real*

Its default attractor is "polished commercial": poreless skin, balanced lighting, tidy set,
locked-off camera. A viewer's brain files that as an ad within one second. A good prompt
**actively pays to get out of that default** — that's a realism layer, not an adjective pile.

Its second attractor is **drift**: past ~10 s with no anchoring, the face slides, the logo
dissolves, the camera loses its logic. The fix is staged structure, not a longer description.

### 2. Never negate a visible object

Seedance handles negation badly on visual content — `no blur` produces blur. But exclusions
work reliably on three other targets:

| Target | Verdict | Example |
|---|---|---|
| A **reference** (what *not* to take from it) | ✅ works, documented | `Do not use the people in @Image 2.` |
| A **default-generated track** | ✅ works | `no background music, no text overlay, no captions` |
| A **manufacturing defect** | ✅ works | `no drift, no deformation, no flickering` |
| A **visible object or attribute** | ❌ backfires | ~~`no red car`~~ → describe what IS there |
| An **image defect** | ❌ backfires | ~~`no blur`~~ → `sharp, in-focus, high clarity` |

> **You never negate a visible object. You negate a source, a track, or a manufacturing defect.**

---

## What it covers

- The official **6-block formula** — Subject + Action + Setting + Style + Camera + Audio
- The **`@` reference system** — assigning an explicit role to each of the 50 materials
- **Staging** for long clips, with physical *and* emotional end-states per beat
- A **7-layer realism pass** so AI UGC reads as a real iPhone video
- **Audio syntax** — music `()`, SFX `<>`, dialogue `{}`, subtitles `【】` — and lip-sync
- **French dialogue**: what's confirmed, what isn't, and the workarounds that hold
- Editing, extension, seamless transitions, blockout / clay render, one-click video
- A **symptom → fix debug matrix** and a delivery checklist

---

## Install

Grab [`seedance-25-prompting.skill`](./seedance-25-prompting.skill) or clone the repo.

**Claude.ai / Claude Desktop** — Settings → Capabilities → Skills → upload the `.skill` file.

**Claude Code** — drop the skill folder in your skills directory:

```bash
git clone https://github.com/gbeyrouti/seedance-prompting-claude-skill.git
cp -R seedance-prompting-claude-skill/seedance-25-prompting ~/.claude/skills/
```

Use `.claude/skills/` instead of `~/.claude/skills/` to scope it to a single project.

**Claude Agent SDK** — point your skills directory at `seedance-25-prompting/`.

Then just ask in plain language: *"write me a Seedance prompt for a 20s UGC ad, here are my
product shots"* — the skill triggers on its own.

---

## What it looks like

**Prompt in:** *"a woman in a Paris café, cinematic, emotional"*

**What most people write — and why it fails:**

```
A cinematic shot of a woman in a Paris café, beautiful lighting, emotional, 8K, masterpiece
```

Nothing here is actionable. No camera logic, no audio track (so the model invents one, usually
a stray music bed), no realism layer — so it renders the default: a coffee commercial.

**What the skill produces instead:** a staged brief with one primary camera move and a speed
word, an explicit audio block, a realism layer with named imperfections, and an exclusions line
targeting tracks and defects rather than objects — plus the recommended aspect ratio, duration
and generation mode underneath.

The full templates live in
[`references/templates.md`](./seedance-25-prompting/references/templates.md).

---

## Model spec

| Parameter | Value |
|---|---|
| Native duration | **30 s in a single pass** (97–721 frames) |
| Extension | 30 s per extension · 60 s nested · Ultra Long Video up to 180 s |
| Output resolution | **480p / 720p** depending on platform. 4K is a marketing claim — check the UI before promising it |
| Aspect ratios | 16:9, 4:3, 1:1, 3:4, 9:16 |
| References | **50 max**: ≤ 30 images (≤ 4K each) + ≤ 10 videos (≤ 30 s total) + ≤ 10 audio (≤ 30 s total) |
| Audio | **native and synchronized** — lip-synced dialogue + SFX + ambience + music in one pass |
| Modes | T2V · I2V · first/last frame · multi-keyframe · storyboard · blockout & clay render · multi-reference · editing · extension · seamless transition · one-click video |
| Prompting language | **English** (or Chinese) |
| API models | `dreamina-seedance-2-5-260628` (BytePlus) · `doubao-seedance-2-5-260628` (Volcano) |

**Real strengths:** long narrative held in one pass, character/product consistency through heavy
references, local editing at a timestamp, green screen, joint audio, 3D blocking.

**Real weaknesses:** complex motion physics, multi-subject interaction, pixel-accurate typography,
fine hand gestures, resolution ceiling. Don't sell Seedance on those.

---

## Repo structure

```
seedance-25-prompting/
├── SKILL.md                          # model spec, golden rule, formula, routing table
└── references/
    ├── structure-et-syntaxe.md       # length, priority, staging, timecodes
    ├── realisme-et-ugc.md            # the 7 anti-default layers, post-production
    ├── references-et-modes.md        # multi-reference, first/last, storyboard, blockout
    ├── editing-et-extension.md       # editing, extension, transitions, one-click
    ├── camera-lumiere-style.md       # camera, lighting, optics, emotion vocabulary
    ├── audio-dialogue-francais.md    # brackets, dialogue, lip-sync, French
    ├── templates.md                  # 7 ready-to-use templates
    ├── debug-et-limites.md           # symptom → fix matrix, limits, checklist
    └── plateformes-et-api.md         # Dreamina, API, fal.ai, model selection
```

Claude reads `SKILL.md`, then loads only the reference file the request actually needs — so a
simple prompt doesn't drag 70 kB of context along with it.

---

## Scope and honesty

- Generated prompts are **in English**. The skill's own documentation is in French.
- The 2.5 prompt grammar is **backward-compatible with 2.0** — the templates work on both.
- The skill explicitly separates what comes from official docs, what comes from field reports,
  and what is extrapolated. Unconfirmed numbers (4K output, extension ceiling, the exact list of
  supported dialogue languages) are **flagged as unconfirmed rather than repeated as fact**.
- The content ecosystem around Seedance is heavily polluted by generated SEO articles. When a
  spec is in doubt, ByteDance / Dreamina / BytePlus documentation beats any third-party comparison.

---

## FAQ

**Does this work with Seedance 2.0?**
Yes. The 2.5 grammar is backward-compatible, so every template runs on 2.0 — you just lose the
2.5-only modes (50 references, 30 s single pass, native audio in one call).

**Do I have to write prompts in French?**
No. The skill's documentation is French; the prompts it outputs are always English, because
that's what the model was trained to parse. You can talk to Claude in any language.

**Can Seedance do French dialogue with lip-sync?**
Partly, and the skill is explicit about where the line is. Announce the language and accent
*before* the line — `Dialogue language: French from France. The woman says in a warm Parisian
accent: {…}` — and check [`audio-dialogue-francais.md`](./seedance-25-prompting/references/audio-dialogue-francais.md)
for what's confirmed versus what's a workaround.

**Why does my character drift after 10 seconds?**
No anchoring. Past ~8 s you need staged structure with an explicit end-state per beat, plus an
identity reference carried across stages. See
[`structure-et-syntaxe.md`](./seedance-25-prompting/references/structure-et-syntaxe.md).

**Why does my AI UGC still look like an ad?**
Because you didn't pay to leave the default. Realism in Seedance is a layer you add on purpose —
[`realisme-et-ugc.md`](./seedance-25-prompting/references/realisme-et-ugc.md) has the 7 of them.

**Seedance vs Veo, Kling, Sora, Hailuo H3?**
Covered in [`plateformes-et-api.md`](./seedance-25-prompting/references/plateformes-et-api.md),
by capability rather than by benchmark score. Short version: Seedance wins on long single-pass
narrative and heavy multi-reference consistency, loses on motion physics and resolution.

**Does it work with Claude Code, Claude Desktop, and the Agent SDK?**
All three — it's a plain Agent Skill (a `SKILL.md` plus reference files). See [Install](#install).

---

## Contributing

Field reports are genuinely useful here, especially on:

- **French** — pronunciation, lip-sync, workarounds that hold or don't
- **Platform gaps** — Dreamina vs API vs third-party relays
- **Confirmed specs** that would replace an uncertainty currently flagged in the skill

Open an issue with the prompt you used, the platform, and what you got back.
See [CONTRIBUTING.md](./CONTRIBUTING.md).

---

## Author

**Grégory Beyrouti** — I test and document AI video models in French on
[YouTube](https://www.youtube.com/@gregory.beyrouti).

If this saved you a few generations, a ⭐ helps other people find it.

## License

MIT — see [LICENSE](./LICENSE).

---

<sub>**Keywords:** Seedance 2.5 prompt guide · Seedance 2.0 prompting · ByteDance video model ·
Dreamina prompt · Jimeng AI video · Doubao Seedance · BytePlus ModelArk · Volcano Ark ·
fal.ai Seedance · Claude Skill · Claude Code skill · Agent Skill · AI video prompt engineering ·
text-to-video · image-to-video · first/last frame · multi-reference video · AI UGC ad prompt ·
lip-sync AI video · native audio video model</sub>

seedance-prompting-claude-skill

seedance-prompting-claude-skill Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

seedance-prompting-claude-skill

seedance-prompting-claude-skill Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README