Your AI video prompts aren't failing because the model is weak. They're failing because you're directing a feature film by describing individual pixels.
That's the core insight behind a quietly impressive GitHub repository that's been accumulating stars since February: Emily2040/seedance-2.0. With 980 stars and 183 forks as of this writing, it's one of the more thoughtfully-engineered AI workflow projects in the video generation space—and the engineering philosophy behind it is worth understanding even if you never use Seedance specifically.
The Model It's Built On
Seedance 2.0 is ByteDance's multimodal video generation system. It debuted at China's 2026 Spring Festival Gala—the first time a publicly-released production extensively used the model for broadcast content—and rolled out globally through CapCut and Dreamina in March 2026.
The technical specs are legitimately impressive: native text, image, video, and audio inputs (up to 9 images, 3 video clips, and 3 audio clips per generation), synchronized audio output in a single pass, 2K resolution at up to 15 seconds per clip, and a Fast variant optimized for low-latency production pipelines. It's available through multiple surfaces including Volcengine's Ark API, Runway's Seedance route, BytePlus ModelArk, and fal.
But here's the thing: a capable model isn't the hard part anymore.
What the Repository Actually Does
The repo's README opens with a declaration: "Direct the model. Don't micro-manage the frame."
That single sentence contains an entire workflow philosophy. Most AI video failures aren't capability failures—they're direction failures. Users arrive with a feeling ("make it cinematic") and attempt to convert that feeling directly into pixel-level instructions, then iterate on adjectives when the output doesn't match their mental image. The repository exists to break that loop.
What Emily2040 has built is effectively an agent operating system—a structured package of skills, references, and evaluation cases that gives any AI assistant (Claude Code, Codex, or other agent platforms) a complete, auditable workflow for Seedance video work. Not a prompt template. An operating system.
The package contains:
- 24 sub-skills covering the full production lifecycle
- 47 reference documents with source-dated platform facts
- 61 evaluation cases testing schema compliance, safety, and output quality
That last point deserves attention: 61 evals. Most AI workflow repos ship a README and some examples. This one ships a test suite.
The Architecture: Three Clusters
The Core Pipeline
The production workflow breaks into specialized sub-skills, each handling a distinct phase:
Brief and Prompt: seedance-interview catches the users who arrive with vague ideas and runs them through a structured creative brief before generating a single token of prompt. The companion seedance-prompt and seedance-prompt-short convert finished briefs into production-ready outputs—full format (~200 words) or compressed (30-100 words for interfaces with token constraints).
Craft Specializations: Separate skills for camera work, motion design, lighting, characters, style, VFX, and audio. Each maps user intent to the specific technical language Seedance responds to. You're not fighting the model—you're speaking its language.
Post-Generation: The seedance-troubleshoot skill is where most users will spend time initially. It doesn't just suggest "try different words." It performs root-cause diagnosis: was the failure in camera direction, reference role assignment, motion specification, lighting setup, or safety wording? Different root causes get different repairs.
The Governance Layer
The three governance skills address the parts of AI video work that usually get skipped until they cause problems.
seedance-copyright handles IP-adjacent requests with a specific approach: it doesn't just refuse. It finds safer creative equivalents—outputs that preserve the emotional and aesthetic intent of the request without touching protected content. That's meaningfully different from a blanket block.
seedance-antislop and seedance-filter work together on the quality/safety axis. The filter specifically handles false positives—benign professional content that gets incorrectly flagged—by providing context that clarifies intent rather than routing around the system.
The Reference Library
The 47 reference documents are what makes this system coherent rather than just comprehensive. A few standouts:
capability-map: Before writing any prompt, you can check what Seedance actually does well versus where it degrades. This prevents spending retake budget on shots the model was never going to nail.
allocation-model: Where does a prompt spend its "fidelity budget"? Identity and character consistency cost differently than motion complexity or scene density. Understanding this changes how you structure prompts.
retake-protocol: A five-verdict triage system for every generated clip: keep, fix in post, edit/extend, re-roll, or rewrite. With a rule that only one variable changes per retake. This is basic scientific reasoning applied to creative iteration—and it's surprising how many workflows skip it.
model-name-map: Source-dated reference preventing Seedance 2.0, Seedance 2.0 Fast, Seedance V2, and various Pro labels from being confused with each other. Mundane but genuinely useful when platform documentation lags behind releases.
The 11-Step Operating Loop
The root SKILL.md specifies how an AI agent processes every video request—an 11-step loop with explicit gates that prevent common failure modes:
- Intake: Extract goal, production phase, target platform, mode, duration, aspect ratio, references, audio needs, and safety/IP flags
- Source gate: Platform claims require loading the dated API status doc before proceeding
- Professional gate: Film, commercial, and delivery work triggers loading the pro-filmmaking standards
- Mode selection: T2V, I2V, V2V, R2V, first/last-frame, edit, extend, or troubleshoot—decided before writing prose
- Capability check: Plan the shot against what the model can actually deliver
- Reference mapping: Assign every asset a primary role (identity, environment, motion, camera, audio, style, etc.)
- Multilingual gate: Chinese, Russian, Japanese, Korean, Spanish inputs get the multilingual vocabulary reference loaded
- Safety gate: IP, likeness, voice, and brand requests route to copyright or filter skills
- Prompt build: Route to the appropriate skill based on what's needed
- Quality pass: Anti-slop check, visible beat, camera move, physical light, audio intent, continuity
- Repair loop: Structured triage after generation, not freestyle adjective addition
Step 2 is worth calling out specifically. The system maintains dated references for platform facts—API availability, model names, pricing, regional access—and requires loading the current-as-of document before making any platform claims. This prevents the common failure mode where AI assistants confidently state outdated information about rapidly-changing platforms.
For Professional Production Teams
The scope extends well beyond individual creators. The package explicitly addresses workflows for:
| Role | Deliverable |
|---|---|
| Director | Treatment, scene beat, coverage, shot endpoint |
| Cinematographer | Shot contract, lens feel, blocking, lighting continuity |
| Producer / Agency | Rights map, approval gates, campaign variants, risk log |
| Editor | Continuity handoff, handles, textless needs |
| Colorist | ACES-aware handoff, HDR/SDR caveats |
| Sound | Dialogue map, stems, M&E, dubbing notes |
| Localization | Subtitles, SDH, forced narratives, market copy |
| Delivery / QC | Frame rate, crop, loudness, naming, human QC checklist |
The multilingual vocabulary layer covers English, Mandarin, Japanese, Korean, Spanish, and Russian—with community-sourced prompt structures for mixed-language workflows, not just mechanical translation.
Installation
For Codex users: `bash python scripts/install_codex_skill.py --force `
For Claude Code or manual installation: `bash git clone https://github.com/Emily2040/seedance-2.0 .agents/skills/seedance-20 `
Once installed, the skill activates on any Seedance-related request. The entry point routes you to the right sub-skill based on where you are in the workflow.
The Bigger Pattern
What's worth watching here isn't just this repository—it's the emerging category it represents.
As generative AI models mature past "impressive demos" into "production infrastructure," the limiting factor shifts from model capability to workflow quality. Seedance 2.0 can generate scenes that would have required a film crew. But generating them well—consistently, safely, at professional quality, within a real production budget—requires the same knowledge that making films with human crews requires.
That knowledge has to live somewhere. This project puts it in a structured, version-controlled, testable artifact that an AI assistant can actually use.
Whether you're building with Seedance or a different model, the template is worth stealing: encode your domain expertise as machine-readable workflow, not as hope that the model will figure it out.
Current version: v5.5.2 (updated June 12, 2026) License: MIT Repository: github.com/Emily2040/seedance-2.0