AGENT SKILL

video-generation

Generate a video clip from the person's own Veo key and compose generated clips, music and captions into a reel.

License 许可
MIT
所属仓库
1aGh/maude
最近更新
2026-11-13
Stars
★ 3
Forks
0

01 前言 Front Matter

name
video-generation
description
Generate a video clip from the person's own Veo key and compose generated clips, music and captions into a reel.

03 SKILL.md 正文

video-generation — the video part of BYOK media generation

Follow host conventions for Claude Code or Codex.

This is the video half of Maude's BYOK media route; images and sound stay in the design plugin. Load Skill design:ai-generation first: it holds the rule you must never break (the agent is never handed a key), the prompt conventions, assets-only localization and the prompt-injection posture, and /design:generate is the one procedure for every modality. This skill adds only what is different for video. Architecture + trust boundary: DDR-164.

Provider

Modality Provider (BYOK, direct) Model Shape Notes
Video Google Veo 3.1 via Gemini API (veo-3.1-generate-preview) async — poll (~1–10 min), job queue native synced audio + image-to-video (--source <still>). maude design generate --modality video --provider gemini. Lands assets/<sha8>.mp4.

Generate a clip

Run /design:generate with --modality video (provider gemini). --source <asset> seeds image-to-video from a still, so a generated clip matches a generated hero's style. The verb waits for the job, and a clip can take up to 10 minutes: pass --timeout, and say so before you send, together with whose account pays (the person's own key) and the licence line below. Paid generation runs only on the person's request or after they confirmed a proposed slot.

Compose into a reel

  • Audio (Phase 2): generated music, SFX and voiceover become Edl.audioTracks[] (layered: duck the music bed under VO with a negative gainDb), and transcribed subtitles become the Edl.captions track. The footage-director codegen renders both (<Audio> + a frame-driven caption overlay). Audio rides the renderMediaOnWeb export path; captions survive the frame-step fallback too.
  • Video (Phase 3): a generated clip is a first-class footage clip. It lands as assets/<sha8>.mp4, and the generate route drops a provenance assets/<sha8>.footage.json stub next to it (the ai-generated tag + the prompt as the summary), so /design-video:reel treats it exactly like ingested footage: the design-video:footage-analyst watches it, then the design-video:footage-director places it as an Edl.beats[].clip. Use it to fill a coverage gap ("no clip for this beat — generate one?"); /design-video:reel Step 1.5 is the flow.

Licence

  • Video (Veo): the provider's terms and any watermark apply. Surface that line before the call; the clip then carries Made by AI in Assets, with the prompt, the provider and the licence line under its details.
  • Never build on OpenAI Sora 2 (API sunset).

04 同仓库更多 Skills