Emily2040/seedance-2.0Public

An agent that reads the scene before it writes the prompt: text, image, video, and reference-to-video for Seedance 2.0, with native audio, IP-safe rewrites, source-dated platform facts, and native reader paths in six languages.

English · 中文 · 日本語 · 한국어

Start here · Skill map · Reference library · Visual gallery · Install

v6.7.0 · MIT · updated 2026-08-01 · what changed

Author: Iamemily2050 (@iamemily2050) · Instagram · X · Website

Surfaces: ByteDance Seedance 2.0 · Dreamina · Jimeng · Doubao · Volcengine Ark · BytePlus ModelArk · Runway Seedance 2 · fal · provider/router surfaces tracked in platform-surface-matrix.md

Most tools ask the model for a "cinematic look." A director asks what the scene is doing — then makes the camera, lens, light, blocking, performance, and sound all serve one intention, in a single recognizable voice, across an entire story.

The directing engine encodes that judgment. It reads a scene's dramatic function — the turn, the point of view, the power, the subtext — names one intention, and derives a coherent setup instead of stacking adjectives.

Ask for "cinematic": epic cinematic shot of a woman reading a letter, emotional, beautiful lighting

Direct it: A woman at a kitchen table reads the letter twice, then her hands lower it and go still. Camera: medium close-up at eye level, a slow push-in that settles when her hands stop. Soft window light keeps her face plain. Sound: room tone, one chair scrape, near-silence — the realization lands in the stilled hands, not a word.

The order is part of the direction: the subject and her action take the opening words — where the model locks in who the shot is about — and the camera, light, and sound follow.

It then holds one directorial voice across every short clip of a long story, and ships with 33 worked derivations — product, music video, horror, anime, action, comedy, documentary, high fashion, sci-fi, and more — each shown end to end.

A reveal is not lit, framed, blocked, or performed like a goodbye. The engine refuses the generic answer and derives the specific one.

Seedance 2.0 Skill OS is English-readable, but the v6 line gives Chinese, Japanese, and Korean readers first-class entry points, active example skills, and native prompt guidance. Keep reference tags exactly as written (@Image1, @Video1, @Audio1, @图片1, @视频1) in every language.

New here? Each language also has a 5-minute quickstart: English · 中文 · 日本語 · 한국어 · Español · Русский.

For longer stories in any language, start with seedance-sequence. For the next part of an accepted clip, use seedance-continuation and update the observed final state before writing the next prompt.

Seedance 2.0 Skill OS is a modular agent-skill package for directing Seedance 2.0 video generations. It is built around a simple principle: direct the model, do not micro-manage the frame.

The repository gives an AI assistant a public, auditable operating system for Seedance work. It defines when to interview, when to write a compact prompt, when to load a technical reference, when to rewrite unsafe IP content, and when to troubleshoot a bad generation.

This skill package turns Seedance 2.0 work into a repeatable assistant workflow:

• Routes vague ideas into short creative interviews instead of premature prompt dumps.
• Directs each scene before drafting: reads its dramatic function, sets one directorial voice, and makes camera, light, blocking, performance, and sound serve a single intention instead of a generic "cinematic" look - and holds that voice across every clip of a long story.
• Writes full or compressed prompts for T2V, I2V, V2V, R2V, FLF2V, edit, extend, audio-aware, and first/last-frame workflows.
• Separates every reference asset by role: identity, environment, motion, camera rhythm, audio tempo, style, or endpoint.
• Keeps model and platform claims source-dated so API, pricing, region, quota, and model-ID details are not guessed.
• Plans into model strengths before drafting: a capability map, a fidelity-allocation model, and a working model of the generator's mechanics that explains why every rule works.
• Runs the shoot like a producer after generation: five-verdict take triage, one-variable retakes, attempt budgets, and cost-aware drafting.
• Provides native-reader front-page paths plus deeper multilingual cinematic vocabulary in English, 中文, 日本語, 한국어, Spanish, and Russian, including role binding, first/last-frame phrasing, edit/extend wording, safety wording, audio cues, continuation wording, and post-production text handling.
• Adds original community-informed examples for Chinese, Japanese, Korean, Russian-English, and Spanish-English prompt structures.
• Adds professional filmmaker workflows for treatment-to-shot-list planning, shot contracts, continuity ledgers, ACES/color handoff, audio post, subtitles/localization, aspect-ratio variants, campaign cutdowns, delivery/QC, and client review packets.
• Handles safe false-positive repairs by clarifying benign production context, not by hiding unsafe intent.
• Rewrites unsafe celebrity, protected IP, private-person, brand, logo, song, or voice requests into safer creative equivalents.
• Diagnoses failed outputs with concrete repair levers: camera, lighting, motion, reference role, duration, framing, audio, or safety wording.
• Ships validation scripts, eval cases, source data, and design checks so maintainers can review changes before release.

Do not blindly ask the skill to extend the original prompt. A continuation must be based on accepted generated footage because Seedance may not end exactly where the original prompt expected.

1. Describe the complete idea and how it ends.
2. The skill divides it into connected clips.
3. Generate Clip 01.
4. Return the generated clip or its final frame.
5. The skill records what actually happened.
6. It writes Clip 02 from the real ending.
7. Repeat until the planned final outcome is reached.

The project state is the source of truth. The clip contract is the current production task. The prompt is a compiled instruction for only that task. Accepted generated footage determines what happens next.

This package is designed for working film and commercial teams, not only for casual prompt writing. It can help an agent produce the artifact the role actually needs:

For these requests, the skill should not stop at a single prompt. It should return the production object first, then the Seedance prompt or prompt batch that fits inside that plan.

Seedance platform behavior changes quickly. Before making factual claims about API availability, face or portrait authorization, upload limits, pricing, regional availability, or model names, load references/api-status.md and check its last_verified date.

This is a Seedance 2.0 skill. A newer line (2.5) launched on consumer surfaces on 2026-07-31 and is out of scope here — the craft below applies to any model, but every platform number is a 2.0 number. Establish which line a surface runs before quoting one. See api-status.md.

As of 2026-06-20, public official sources describe Seedance 2.0 as supporting text, image, audio, and video inputs. Official launch and model-card material says references can include up to 9 images, 3 video clips, and 3 audio clips.

Volcengine and BytePlus docs now expose Seedance 2.0 Mini as a surface-specific model lane. Treat Seedance V2 Mini as shorthand for Seedance 2.0 Mini only when the active surface confirms it. Current source-visible IDs include doubao-seedance-2-0-mini-260615 on Volcengine and dreamina-seedance-2-0-mini-260615 on BytePlus.

Volcengine docs also keep doubao-seedance-2-0-260128 and doubao-seedance-2-0-fast-260128 visible as Ark model IDs and document first/last-frame role usage on that surface. Runway documents seedance2 with 5-15 second duration and optional image, video, and audio references.

Third-party provider/router pages tracked as of 2026-06-20 include EvoLink, OpenRouter, Kie.ai, PiAPI, LaoZhang, Runware, ModelsLab, AI/ML API, MuAPI, SeeGen, and Segmind.

Treat every endpoint, model ID, price, account requirement, face/reference policy, and output-rights claim as provider-specific and recheck live before implementation. China-facing searches should prefer official ByteDance, Volcengine, BytePlus, Doubao, Jimeng/Jianying, and CapCut/Jianying surfaces; workflow hosts or business-partner news are not public API providers unless they publish provider-owned API docs.

Access, pricing, upload limits, regions, resolution, audio-combination rules, and authorization requirements remain surface-specific.

The v6 release line keeps a dated research layer for safer data mining, multilingual prompting, sequence-state work, and platform claims:

• research-2026-05-30.md records official and field-observed signals.
• platform-surface-matrix.md separates model capability from Dreamina/Jimeng, Volcengine/Ark, BytePlus, ComfyUI, and provider/router behavior.
• model-name-map.md prevents Seedance 2.0, Seedance 2.0 Fast, Seedance 2.0 Mini, Seedance V2, and ambiguous Pro labels from being mixed together.
• community-source-methodology.md explains how to mine public prompt corpora without copying unsafe examples.
• multilingual-community-examples.md captures safe mixed-language and localized prompt structures from community pattern mining.
• pro-filmmaking-standards.md adds industry workflow boundaries for shot lists, continuity, color, audio, localization, and delivery.

The diagram is the contract: every request passes the gates, the root routes it, and the validators hold the line. Six lanes stay separate by design:

• Research sources: dated official, academic, platform, and community evidence.
• Production spine: brief, shot list, continuity, post handoff, localization, and delivery/QC.
• Prompt router: interview, prompt writing, compression, recipes, and troubleshooting.
• Multimodal references: image, video, audio, first-frame, last-frame, and role-bound assets.
• Safety gates: IP, likeness, voice, brand, real-person, filter, and platform-policy checks.
• Quality evals: schema checks, source freshness, vocabulary integrity, design audit, and behavior cases.

Concept art for the system, generated and curated. Every image is paired with searchable alt text so the gallery stays auditable; the README's working visuals above are hand-built vector assets that follow the design standard.

Client support for Agent Skills is still tool-specific. Codex documents a skill as a directory with a required SKILL.md, optional scripts/, references/, assets/, and optional agents/ metadata.

Codex scans .agents/skills locations from the working directory upward, plus user/admin/system skill locations. A repository root with SKILL.md is shaped like a skill folder, but it still needs to be installed/copied under a scanned skills directory or distributed as a plugin for automatic discovery.

Every install path below runs from inside a local copy of this repository, so start here:

git clone https://github.com/Emily2040/seedance-2.0.git
cd seedance-2.0

Without git installed, use the green Code → Download ZIP button on the repository page, unzip it, and change into the unzipped folder instead. Nothing else on this page works until one of those two has happened.

The installer is not Codex-only. It copies the skill into any client that reads a skills directory — point --dest at the directory yours scans:

# Codex (default: $CODEX_HOME/skills, else ~/.codex/skills)
python scripts/install_codex_skill.py

# Claude Code — personal install, available in every project
python scripts/install_codex_skill.py --dest ~/.claude/skills

# Any client — install into another project, run from that project
python /path/to/seedance-2.0/scripts/install_codex_skill.py --dest .claude/skills

The command copies the repository to <dest>/seedance-20 and prints where it landed. Add --force only to replace an install that already exists — it deletes the previous copy first, so it is not the flag to start with. Restart your client afterwards so seedance-20 appears in its skill list.

Installs skip the image gallery (about 18 MB of PNGs), the test suite, and the network-capable evaluator — an agent needs the skill text, not the artwork. The gallery links in an installed copy's README therefore resolve only in this repository.

A destination inside this repository is refused rather than attempted: copying the source tree into a directory inside itself recurses until the path length fails. For a project-local install, run the script from the project you are installing into, by absolute path, as above.

This repository keeps dense facts in references so the active skill stays small.

If your client supports installing a skill directly from a GitHub repository, use this repository URL:

https://github.com/Emily2040/seedance-2.0

For manual installation, copy this repository into the skill directory used by your agent client. The directory name should match the root skill name, seedance-20. Treat the table below as common local targets to verify in your own client, not a universal support guarantee.

Several of these clients share the .agents/skills/ convention — Codex, Google Antigravity, OpenCode, Amp, and Goose all read it — so one install under .agents/skills/seedance-20/ can serve them together, and .claude/skills/ is read by many as a compatibility path. Install once as the seedance-20 root skill; its sub-skills and references resolve by relative path.

Run these checks before every release. The source-registry check runs with --enforce-freshness here so a stale registry blocks a release; per-pull-request validation deliberately omits the flag, because staleness depends on the calendar rather than on the change under test.

Every check is offline. Exactly one of them, schema_check.py, needs a third-party library, so install the hash-pinned lock first — on a clean checkout that step otherwise stops the run:

python -m pip install --require-hashes --requirement requirements-validation.lock

python scripts/validate_skills.py --strict
python scripts/content_audit.py --strict
python scripts/eval_schema_check.py --strict
python scripts/schema_check.py --strict
python scripts/design_audit.py --strict
python scripts/build_hero.py --check
python scripts/source_registry_check.py --strict --enforce-freshness
python scripts/vocab_schema_check.py --strict
python scripts/project_state_check.py --strict
python scripts/continuity_chain_check.py --strict
python scripts/behavior_contract_check.py --strict
python scripts/sequence_eval_check.py --strict
python scripts/generation_run_check.py --strict
python scripts/prompt_architecture_stress.py --strict
python scripts/prompt_lint.py --self-test --strict
python scripts/eval_run.py --self-test --strict
python scripts/extract_last_frame.py --self-test --strict
python -m unittest discover -s tests -v
python -m compileall scripts tests
git diff --check

The CI workflow runs this same list on push and pull request, with the one deliberate difference noted above: it omits --enforce-freshness. These checks are deterministic and offline — they prove the package is well-formed.

Whether references/source-registry.md is stale depends on today's date, not on the change being tested, so it is not asked per pull request — that would fail unrelated work on a calendar boundary. It is asked in two places instead:

Drift and staleness are tracked in a single automatically maintained issue. It opens when the registry first drifts, is refreshed in place each week rather than re-notifying, and closes itself once the registry is back inside the window.

The scheduled job never edits the registry. Re-stamping last_verified without actually re-reading the upstream sources would record a verification that never happened, so refreshing it is deliberately a human step.

To prove the package is also good, run the model-in-the-loop harness, which sends each eval case through the real skill content and scores the response against the case's assertions using eval-rubric.md:

export ANTHROPIC_API_KEY=...   # required for a live scored pass
python scripts/eval_run.py --run --ledger evals/eval-run-ledger.md --stamp 2026-06-28

This is the quality gate, not a shape gate, so it lives outside offline CI; the latest scored run is recorded in evals/eval-run-ledger.md.

The front page follows an editorial design system rather than default AI styling: warm ink and paper themes, a serif display face paired with monospace specification labels, a single amber accent, and hairline rules — no gradients, no glow. Camera motifs are deliberately retired: no viewfinder marks, timecode, record dots, or aspect badges. They depict the tool rather than the work, and they are the visual cliché of every AI-video product.

The masthead is generated from one geometry by scripts/build_hero.py, so its dark and light variants cannot drift apart; --check proves the committed SVGs still match the generator, and it runs in CI.

The masthead and the hand-built operating diagram (assets/hero-dark.svg, assets/hero-light.svg, assets/skill-map.svg) are served through a prefers-color-scheme picture element; generated bitmap art lives only in the curated visual gallery, including the text-rich infographics.

The README must stay readable in GitHub mobile, dark mode, and narrow widths. SVG assets must include <title> and <desc> elements, use internal CSS only, and avoid external fonts, scripts, or resources. Tokens and rules live in references/frontend-design-system.md and docs/frontend-redesign.md.

See CHANGELOG.md. Current release: v6.7.0.

MIT © 2026 Iamemily2050 (@iamemily2050)