Separation of creative direction and technical instructions
Seedance 2.0 Skill OS is an agent operating system for directing AI video generation, primarily written in Python, licensed under MIT and created on 2026/2/25. Its core design is intent-first direction; the director states an intention, the system routes it to role specialists (cinematographer, editor, colorist), and the generator receives focused role instructions. It separates creative direction from technical instructions, keeping the director focused on creative requirements while technical prompts are handled separately.
It is built on ByteDance's Seedance 2.0 video generation model; Seedance 2.0 supports text, image, audio and video inputs, with reference limits of 9 images, 3 video clips and 3 audio clips; variants include Seedance 2.0 Mini, Seedance 2.0 Fast and Runway's seedance2 (5–15 seconds of output). Skill OS's four input modalities are text prompts, image references, video references and audio references, with native audio support added in v6. Each reference has a defined role, such as identity, action, environment, camera rhythm, style or endpoint.
The repository contains 28 subskills, 56 reference documents and 112 evaluation cases, including 14 core pipeline skills. It is designed for working video and advertising teams, with shot contracts, continuity ledgers, ACES and color-grading delivery workflows; it is unsuitable for casual users. The current version is v6.0.1 (2026-06-20), focused on CJK native-reader expansion and the v6 audit.
interview
sequence
Generate Clip 01
Return the ending frame
continuation
Repeat until complete
Project state is the single source of truth. The clip contract is the current production task, and the prompt is an instruction compiled only for that task. The accepted generated video determines what happens next.
Agent client skill pack installation
Seedance 2.0 Skill OS runs as a skill pack; install it by copying the repository into your agent client's skill directory. The official project provides a single-command Codex installation script and supports manual copying for other clients. It supports Claude Code, Codex, OpenClaw and Hermes; clone the repository first, then select the installation method for your client.
# Get the repo
git clone https://github.com/Emily2040/seedance-2.0.git
cd seedance-2.0
# Codex one-step installation: copy to $CODEX_HOME/skills/seedance-20 or ~/.codex/skills/seedance-20
python scripts/install_codex_skill.py --forceManual installation in the client skill directory
Other clients use manual installation; copy the repository into the client's skill directory, with paths varying by platform. For example, Claude Code uses .claude/skills/seedance-20/, while Codex uses .agents/skills/seedance-20/.
# Claude Code
cp -r seedance-2.0 ~/.claude/skills/seedance-20
# Agent runtimes such as Codex
cp -r seedance-2.0 ~/.agents/skills/seedance-20Core pipeline skills and role assignments
Skill OS divides production into skills with defined inputs and outputs. The eight most commonly used core pipeline skills are listed here. Production usually starts with seedance-interview to turn a vague idea into a creative brief, followed by seedance-sequence to create clip contracts, then the relevant domain skills for cinematography, camera movement, lighting and other dimensions.
Brief · 01
seedance-interview
Creative interview
Take a vague idea and return a focused creative brief and next-step path. Turn "I want to make something" into an actionable direction.
Plan · 02
seedance-sequence
Shot planning
Take a long story or multi-clip request and return a story spine, continuity bible, Clip 01 contract and Clip 01 prompt. Produce only the first clip at a time.
Plan · 03
seedance-continuation
Next-clip continuation
Take the generated video and next beat, record the observed state and write the next clip from the actual ending rather than infer it from the original prompt.
Prompt · 04
seedance-prompt
Full prompt
Take a clear concept and return a complete prompt. Compile an established creative intention into generator-ready instructions.
Prompt · 05
seedance-prompt-short
Compressed prompt
Compress a prompt to 30–100 characters for providers that require concise input or are sensitive to length.
Domain · 06
seedance-camera
Camera direction
A domain skill for camera movement. Provide focused descriptions of camera motion for the generator to execute.
Domain · 07
seedance-motion
Motion direction
A domain skill for on-screen movement. Describe subject and scene movement to turn intent into specific motion instructions.
Domain · 08
seedance-lighting
Lighting direction
A domain skill for lighting. Turn atmosphere and lighting intentions into lighting descriptions the generator can understand.
Three governance skills for safety and quality checks
Beyond pipeline skills, Skill OS has three governance skills for rewriting risky content, removing prompt bloat and correcting false positives. Use them to check risks and quality before sending output to the generator.
| Governance skill | What it handles | Principle |
|---|---|---|
seedance-copyright | Prompt rewriting involving IP, celebrities and brands | Reduce infringement risk |
seedance-antislop | Prompt bloat and redundant words | Keep prompts focused |
seedance-filter | Correcting false positives | Clarify benign context; never conceal intent |
Eight operating points
These operating points come from Seedance 2.0 Skill OS's README and references. Most misuse centers on recurring questions about which skill to start with, what to read for continuation and how to verify provider settings.
seedance-interview for vague ideas
Do not write a prompt immediately. Start any vague idea with seedance-interview, which returns a focused creative brief and next-step path. Clarify the direction before production.
Source · Official repo / README
Actual endings and continuation evidence
seedance-continuation writes the next clip from the actual ending of accepted generated video because Seedance may stop somewhere other than the original prompt expected. Continuation must use the actual final frame.
Source · Official repo / README
seedance-prompt-short concise output
When a provider is sensitive to length or you need concise input, use seedance-prompt-short to compress the prompt to 30–100 characters.
Source · Official repo / README
seedance-antislop redundancy removal
Prompt bloat dilutes intent. seedance-antislop removes redundant words and prompt bloat, focusing input on descriptions that affect output.
Source · Official repo / README
seedance-copyright IP / brand risk rewriting
Run seedance-copyright before generation when celebrities, brands or protected IP are involved. It rewrites relevant prompts to reduce infringement risk before they reach the generator.
Source · Official repo / README
Multilingual skills and native-language support
Skill OS includes 9 vocabulary skills, with native support for English, Chinese, Japanese, Korean, Spanish and Russian. Chinese, Japanese and Korean have corresponding example skills, seedance-examples-zh / -ja / -ko.
Source · Official repo / README
Project state as the factual basis
During production, use project state as the authority; the clip contract is the current task, the prompt is compiled only for that task, and accepted generated video determines the next step. Do not continue from memory or old prompts.
Source · Official repo / README
Pre-launch provider model ID checks
Endpoints, model IDs, prices and policies vary by surface. references/api-status.md records the last verification date (as of 2026-06-20); check it against current provider information before implementation.
Source · Official references/api-status.md
Vague ideas and second-clip generation
In this illustrative workflow, seedance-interview turns a vague idea into a brief, seedance-sequence returns the Clip 01 contract and prompt, and Clip 01 generation returns the actual ending. Then seedance-continuation writes Clip 02 from the accepted ending. Skill names come from the official repository; the content is illustrative.
$ You › seedance-interview
I want to make a short film set in a city on a rainy night, with a film noir atmosphere, in about three segments.
claude: seedance-interview ›
Clarify the intent before writing any prompt:
hl: Who is the protagonist? Does the camera follow them or stay fixed? Is the rain atmosphere or part of the plot?
$ You › seedance-sequence
claude: seedance-sequence ›
Clip 01:A lone person walks across wet pavement reflecting neon lights, followed by a low-angle camera.
hl: [Output only the prompt for Clip 01; continue subsequent clips later]
ok: accepted · ending:the character stops in front of a lit doorway instead of continuing forward
$ You › seedance-continuation
claude: seedance-continuation ›
Record the observed state: the character has stopped and faces the lit doorway.
hl: Continue from the actual ending(not the original prompt):
Clip 02:He pushes the door open; the warm indoor light contrasts with the cold rain outside.
Clip 01 ends in front of a lit doorway, not at the location expected by the prompt. Continuation proceeds from this actual ending instead of inferring it from the original prompt.
Project state and continuation conditions
Intent-first direction uses state as the sole source of truth and does not write the entire video's prompt at once. seedance-sequence produces only Clip 01 because the generator's actual output cannot be determined beforehand. After Clip 01 is generated, seedance-continuation records the actual ending, then writes Clip 02 from it.
Continuation must use accepted generated video. Seedance may stop somewhere other than the original prompt expected; continuing from the original plan can break continuity. The director states the intention, while the system compiles it into focused instructions for each clip.
Capability boundaries and usage limits
Skill OS production workflow integration
After validating the basic intent-first workflow, integrate Skill OS into video production. It covers multilingual examples, multimodal workflows and current provider verification practices.
Advanced configuration directions
1. Use multilingual example skills for native-language context. Chinese, Japanese and Korean have corresponding seedance-examples-zh / -ja / -ko skills. Combine them with 9 vocabulary skills to make prompts fit non-English contexts. CJK native-reader expansion is the theme of v6.0.1.
2. Check examples-by-mode for the relevant mode. references/examples-by-mode.md organizes examples by T2V, I2V, V2V, R2V, FLF2V, edit, extend and other modes. Identify your scenario's mode before selecting a reference.
3. Keep api-status verification current. Before implementation, check the last verification date for endpoints and model IDs in references/api-status.md (as of 2026-06-20). Provider surfaces change frequently; current checks help avoid invalid IDs.
4. Build a continuity bible and shot contracts for professional delivery. For multi-clip production, use the story spine and continuity bible from seedance-sequence as the main record and advance through clip contracts one at a time. This workflow supports video / advertising team delivery.
5. Check governance skills before generation. Before sending content to the generator, use seedance-copyright, seedance-antislop and seedance-filter as needed to maintain quality and compliance during scaled generation.
Three reference documents
① references/api-status.md: Last verification records for endpoints and model IDs; check current information before implementation.
② references/examples-by-mode.md: Examples by T2V / I2V / V2V / R2V / FLF2V / edit / extend mode.
③ references/multilingual-community-examples.md: Multilingual community examples for non-English contexts.
The director states the intent, the system routes it to a role specialist, and the generator receives focused instructions for that role. Separating creative direction from technical instructions is central to this design.