SuperAPI AI Directory

video-generation

Agents & models

View repository

Updated 10-15 16:41

About this skill

Video Generation

Use When

  • The desired final asset is a generated video or video render plan.
  • The request may be a simple clip, storyboard, UGC, podcast clip, reference video, talking-head, image-to-video, text-to-video, or research handoff.
  • The next decision is task class, request architecture, and execution handoff.

Do Not Use When

  • The user only needs transcription, subtitles, semantic analysis, or edit decisioning. Use media-router.
  • The request architecture is already approved and provider-ready. Use seedance-submitter or video-batch-runner.
  • The task is image-only or audio-only.

Core Boundary

This is the video generation controller. It does not submit jobs.

It must:

  1. classify the video task,
  2. choose whether references or storyboard must be resolved,
  3. hand approved structure to video-request-architect,
  4. hand submit-ready plans to seedance-submitter or video-batch-runner.

Task Classes

Task class Use when Required handoff
simple_clip one short text/image-to-video clip video-request-architect
storyboard panel sequence or contact sheet drives the clip storyboard-grid-writer, then video-request-architect
ugc creator-style product or app video ugc-flow unless it is one isolated clip
podcast voice-led clip, audiogram, or speaker segment audio-generation, then video architecture
reference_video benchmark video controls rhythm or motion reference-decode and reference-contract-builder
research_handoff research/benchmark becomes a generation request benchmark-to-brief, then video-request-architect

Routing Table

If not video-generation Send to
Understanding or subtitles first media-router
Image asset must be created first image-generation
Voice or dub must be created first audio-generation
Product-to-UGC pipeline ugc-flow
Provider-ready request exists seedance-submitter or video-batch-runner

Three-Layer Boundary

  • Controller: classify task and produce the handoff. It must not submit.
  • Architect: video-request-architect builds provider-agnostic structure. It must not call providers.
  • Runner: seedance-submitter or video-batch-runner validates and executes normalized requests. It must not make creative strategy decisions.

Output Shape

Return:

  • taskClass
  • requiredUpstreamSkills
  • architectureHandoff
  • runnerHandoff
  • referencePolicy
  • mustNotDo

Stop Conditions

  • Stop when required user intent, source evidence, or owned input artifacts are missing and guessing would change the result.
  • Do not ask a runner to infer hook, persona, or reference meaning.
  • If an owned CLI or script command fails, report the exact error and stop. Do not bypass the failure with metadata-only answers, readiness probing, local payload rewrites, fallback providers, or unpublished tools.

Public Command Boundary

  • Choose the smallest matching command or workflow from the user input and run it directly.
  • If an owned CLI or script command fails, report the exact error and stop. Do not bypass the failure with metadata-only answers, readiness probing, local payload rewrites, fallback providers, or unpublished tools.
  • This public skill is instruction-driven. Produce the controller handoff artifact directly from the available evidence.
  • Do not call private provider/runtime paths or unpublished local tools.
  • If the CLI returns a quote-confirmation challenge, run postplus quote confirm --json --challenge-file <challenge.json> and retry with the returned token.