seedance-2.5
Skill3 starsChanged 3 months agoMITView on GitHub
Compose Seedance 2.5 video generation prompts from a VideoProjectSpec and compile them into a generation-request with this repo's compiler. Use for ByteDance Seedance 2.5 text-to-video, image-to-video, video-extension, and video-edit jobs: a 30-second-per-generation audio-video joint generation model that supports precise reference control, white-model control, green-screen editing, professional camera movement and performance blocking.
In this repository3 other files it gives its agents.
- llms.txtllms.txt
- seedance-minimax-promptsSkill
- minimax-h3Skill
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. About 0 are common, 10 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈10
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 18
Use this skill when the deliverable targets ByteDance Seedance 2.5 and the job is one of:
- Very specific: line 27
Do not use it for other models (MiniMax H3 has its own skill), for pure image generation, or for content with unlicensed/movie/TV/brand …
- Distinctive: line 32
Ask for or extract, in this order:
- Distinctive: line 46
If the user wants more than 30 s, plan the first 30 s now and note that an extension (up to two, official limit) will be a separate …
- Distinctive: line 77
Use this repo's compiler — do not hand-roll the prompt:
- Distinctive: line 97
Write audio intent (dialogue / sound design / music) into every prompt.
- Distinctive: line 98
Write the duration explicitly.
- Very specific: line 103
Keep all examples original.
- Very specific: line 107
Don't write duration_seconds > 30, even with an extension flag.
- Distinctive: line 108
Don't claim extension beyond the official twice limit.
- Distinctive: line 109
Don't let two references define the same identity or style.
- Distinctive: line 110
Don't copy third-party prompt collections into this repo.
- Distinctive: line 111
Don't present third-party API features as ByteDance official capabilities (see hard rule below).
- Very specific: line 132
Not on the official page (community practice / product-surface, must be marked as such, never official): first-frame/last-frame mode …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Seedance 2.5 — Video Prompt Skill
Turn a video brief into a validated Seedance 2.5 generation request.
When to use
Use this skill when the deliverable targets ByteDance Seedance 2.5 and the job is one of:
- text-to-video — a new scene from a written brief only;
- image-to-video — one or more stills anchor appearance and composition;
- video-extension — continue an existing clip (≤ 30 s per generation, extend at most twice);
- video-edit — edit an existing clip, including audio editing requests.
Do not use it for other models (MiniMax H3 has its own skill), for pure image generation, or for content with unlicensed/movie/TV/brand assets.
How to collect requirements
Ask for or extract, in this order:
- Mode — text-to-video / image-to-video / video-extension / video-edit.
- Duration — number of seconds, 1–30 (hard maximum per generation).
- Subject — the one who/what the video is about, with stable attributes.
- Scene — setting, lighting, mood.
- Action beats — ordered micro-changes that make up the 30-second story.
- Camera — one dominant move per shot, shot size, angle, lens character.
- Audio intent — dialogue, sound design, music. Always; Seedance 2.5 is an audio-video joint generation model.
- References — which assets exist and what single job each one has (first frame, identity, camera grammar, motion, music).
- Constraints — fixed requirements and negatives (no captions, no text).
If the user wants more than 30 s, plan the first 30 s now and note that an extension (up to two, official limit) will be a separate generation.
How to build a VideoProjectSpec
Build a spec dict that satisfies schemas/video-project.schema.json. Required
fields: spec_version, project_goal, model (seedance-2.5),
duration_seconds (1–30), aspect_ratio, subject (non-empty string),
action, evidence. mode is optional — the compiler infers it from the
reference assets (video → video-extension/video-edit, image →
image-to-video, none → text-to-video) — but set it explicitly when the
intent must not be guessed. Recommended fields: scene
(setting/lighting/mood), camera (movement/angle/lens),
audio (dialogue/sound_design/music), continuity, constraints
(flat list or preserve/modify/forbidden block), resolution
(aspect_ratio, width, height). Optional: references.images/videos/audio,
first_frame, last_frame, performance.direction, extension.count (0–2,
only in video-extension mode), provider.
Mode-specific requirements:
image-to-video: ≥ 1 reference image or afirst_frame/last_framedescription.video-extension: ≥ 1 reference video andextension.countin 0–2.video-edit: ≥ 1 reference video.
Reference manifests may describe hypothetical original assets; never reference unlicensed material.
How to compile the final prompt
Use this repo's compiler — do not hand-roll the prompt:
from ai_video_director.compilers.seedance_25 import compile
request = compile(spec, manifest) # manifest optional
prompt = request["prompt"] # single well-structured prompt stringThe compiler validates constraints (duration ≤ 30 s, extension twice at most,
mode known, subject present, reference requirements per mode) and raises
ai_video_director.validators.ValidationError on violations. It returns a
generation-request-shaped dict: model, mode, prompt,
reference_manifest, duration_seconds, resolution_hints, provider,
evidence. The prompt follows models/seedance-2.5/prompt-grammar.md.
Do / Don't
Do:
- Write audio intent (dialogue / sound design / music) into every prompt.
- Write the duration explicitly; keep it ≤ 30 s per generation.
- Give every reference exactly one job; distinguish "edit this clip" from "reference this clip".
- Repeat stable identity/wardrobe/setting details in the continuity line.
- Name camera moves and performance blocking explicitly.
- Keep all examples original; no movie/TV/brand IP, no celebrities.
Don't:
- Don't write
duration_seconds> 30, even with an extension flag. - Don't claim extension beyond the official twice limit.
- Don't let two references define the same identity or style.
- Don't copy third-party prompt collections into this repo.
- Don't present third-party API features as ByteDance official capabilities (see hard rule below).
Official capability facts (checked 2026-08-19)
Source: https://seed.bytedance.com/en/seedance2_5 — checked_at: 2026-08-19.
| Fact | Wording on the official page |
| Audio-video joint generation | "Seedance 2.5 is a next-generation audio-video joint generation model". |
| 30 s per generation | "Create videos up to 30 seconds in a single generation". |
| Extend twice | "with the option to extend twice for richer, more complete storytelling" / "supports two video extensions". |
| Motion & realism | "Motion is smoother and more consistent, and the visuals are more realistic." |
| Precise reference control | "built for 30-second storytelling with precise reference control". |
| Reference video understanding | "Understands reference videos more precisely -- capturing the intention, framing, and cinematic language to go beyond motion transfer into creative interpretation." |
| Reliable editing | "Editing becomes more reliable, and responds to a wider range of audio and visual editing requests." |
| White-model control | "Provides advanced capabilities such as white-model control". |
| Green-screen editing | "and green-screen editing". |
| Professional camera movement | "combined with professional camera movement". |
| Performance blocking | "and performance blocking, to better support complex video production". |
Not on the official page (community practice / product-surface, must be marked
as such, never official): first-frame/last-frame mode mechanics, @Image1 /
@Video1 / @Audio1 inline reference syntax, reference-count limits,
aspect-ratio lists, beat-planning timing ranges, drift-vs-duration heuristics.
Hard rule — third-party API features
Third-party API features — native 4K, 72 API routes, relaxed moderation, specific pricing, endpoints — must never be presented as ByteDance official capabilities. The official page checked on 2026-08-19 does not state them. If a third-party guide or API provider claims them, attribute them to that provider explicitly or omit them; a mismatch must be reported, not silently adopted.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

