seedance-minimax-prompts
Skill3 starsChanged 3 months agoMITView on GitHub
Router for AI video generation prompt skills in the seedance-minimax-prompts repository. Use when the user wants video generation prompts or generation requests for ByteDance Seedance 2.5 (text-to-video, image-to-video, video-extension, video-edit) or MiniMax H3 / Hailuo (text-to-video, first-frame, last-frame, first-last-frame, Ref2VA). Routes to the per-model skills, explains the unified VideoProjectSpec -> route() -> compiler flow, and enforces hard facts: MiniMax H3 duration 4-15 s, Seedance 2.5 up to 30 s per generation with at most two extensions, third-party claims never presented as official, no claim of an official SDK, a complete local open-source H3 2K pipeline, or native 4K Seedance.
In this repository3 other files it gives its agents.
- llms.txtllms.txt
- minimax-h3Skill
- seedance-2.5Skill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 24
Use this skill when the user asks for video generation prompts, prompt writing, or generation requests targeting one of these models:
- Very specific: line 32
Do not use it for other models, for pure image generation, or for content with unlicensed/movie/TV/brand assets.
- Distinctive: line 44
Read the model's grammar before writing a prompt:
- Very specific: line 51
Do not hand-roll prompts. Build a VideoProjectSpec dict conforming to schemas/video-project.schema.json, then route it:
≈ “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
AI Video Director Skills — Router
This is the entry-point skill of the seedance-minimax-prompts repository:
a neutral, open-source director layer for AI video generation prompts. It is not
an official ByteDance or MiniMax product, not an SDK, and not a model hosting
service.
When to use
Use this skill when the user asks for video generation prompts, prompt writing, or generation requests targeting one of these models:
- Seedance 2.5 (ByteDance) — modes:
text-to-video,image-to-video,video-extension,video-edit. - MiniMax H3 (Hailuo) — modes:
text-to-video,first-frame,last-frame,first-last-frame,ref2va.
Do not use it for other models, for pure image generation, or for content with unlicensed/movie/TV/brand assets.
Routing
- Determine the target model from the user's request or from
spec["model"]("seedance-2.5"or"minimax-h3"). Load the matching sub-skill and follow it:
skills/seedance-2.5/SKILL.md(skill name:seedance-2.5)skills/minimax-h3/SKILL.md(skill name:minimax-h3)
Read the model's grammar before writing a prompt:
models/seedance-2.5/prompt-grammar.mdmodels/minimax-h3/prompt-grammar.md
How to use the compiler
Do not hand-roll prompts. Build a VideoProjectSpec dict conforming to
schemas/video-project.schema.json, then route it:
from ai_video_director.router import route
request = route(spec) # validates spec + per-model constraints
prompt = request["prompt"] # model-specific, grammar-compliant promptFlow: validate_spec() (schema + hard constraints) -> route() (pick compiler by
spec["model"]) -> model-specific compile() -> validate_generation_request().
Compilers live in src/ai_video_director/compilers/seedance_25.py and
src/ai_video_director/compilers/minimax_h3.py. Violations raise
ai_video_director.validators.ValidationError; nothing is silently clamped.
Hard rules
- Duration caps are hard. MiniMax H3: 4-15 s only (open weights and API).
Seedance 2.5: up to 30 s per single generation; longer timelines use the
official
video-extensionflow with at most two extensions, each extension its own generation within the cap. Never clamp, round, or silently adjust a duration. - Never present third-party claims as official. Claims that appear only in third-party wrappers — native 4K Seedance, 72 API routes, relaxed moderation, specific pricing or endpoints — must be attributed to that provider or omitted.
- Never claim: an official SDK (official surfaces are product pages, the API and the open-source repo), a complete local open-source H3 2K pipeline (H3-Context-IR and H3-Regenerate-2K are not open-sourced; local H3-Base outputs 768p, 2K requires the official API), or native 4K Seedance output.
- Ref2VA limits are hard: up to 9 images, 3 videos, 3 audio, 12 files total; each reference video/audio clip 2-15 s, total reference duration <= 15 s.
- No model capability inheritance. A mode or limit from one model is never
assumed valid for the other. Model modes are per-model: Seedance 2.5
text-to-video/image-to-video/video-extension/video-edit; MiniMax H3text-to-video/first-frame/last-frame/first-last-frame/ref2va. - Audio intent is always written into the prompt: both models are audio-video joint generation (Seedance 2.5 official; H3 native 32 kHz stereo).
Where evidence lives
- Capability facts:
models/<model-id>/capability.yamlandconstraints.yaml(every entry carriesevidence_level,checked_at,source_url,surface). - Source registry:
sources/official-sources.yaml(official only) andsources/third-party-providers.yaml(third-party, kept separate). - Human-readable registry and licensing verdict:
THIRD_PARTY_SOURCES.mdanddocs/license-audit.md. - Facts checked on 2026-08-19; when re-verifying, update
checked_atandsource_urltogether.
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.

