Lohnt sich dein eigenes Revid AI?

Build me a Claude Code skill that makes short-form videos to replace Revid AI. Requirements:

- The deliverable is a skill folder, not a web app: SKILL.md plus a Node + TypeScript CLI
  run with tsx, no build step. ffmpeg does every stitch, mux, and caption burn-in.
- Projects live in ./projects/<slug>/ with brief.md, script.md, shots.json, and a render/
  folder, so everything about one video sits in one directory I can delete.
- Styles are files, not prompts I retype: styles/<name>.json holds the look, aspect ratio,
  pacing, caption font and position, voice id, and a shot-prompt preamble. A project names
  one style and inherits all of it.
- Generate each shot with Seedance 2.0 when it needs native audio or lip sync, or Gemini
  Omni Flash when I want to revise it conversationally without re-prompting · the choice
  is per shot in shots.json. Keys in .env.
- Narrate with ElevenLabs where a shot has voiceover rather than native dialogue, and burn
  captions from the returned word timings, styled by the style file.
- Cache every clip by a hash of its prompt plus style, so re-rendering a project never
  pays twice for a shot that did not change.
- SKILL.md must tell the agent to read the live API docs before writing any call:
  replicate.com/bytedance/seedance-2.0, ai.google.dev/gemini-api/docs/video, and
  elevenlabs.io/docs. These models ship fast, so never write a request shape from memory.
- No accounts, no telemetry, everything on my machine except the three model APIs.
  Secrets in .env with a committed .env.example.
- Out of scope: publishing to TikTok, Instagram, or YouTube, and any stock or viral-clip
  library. Do not build a web UI, the interface is the skill and the CLI.
- README: the .env keys, how to install ffmpeg, and a worked cost estimate for a 30 second
  video at current per-second model prices, so I know what a render costs before I run it.