Auto-render: script → spec → video file¶

How to let an AI agent produce finished short videos for the Auto formats in formats.md without an editor. Any code-based video tool works; the example uses Remotion (React → MP4; free for individuals and small companies, check its license for larger teams).

The pipeline¶

post file (marketing/<platform>/NN-name.md)
  ## Title · ## Script · ## Caption · ## Production · ## Render  ← the agent writes all five
        │
        ▼  render script: extract the ## Render JSON → props.json
  video tool (one "Spec" composition that reads props)
        │
        ▼
  marketing/<platform>/out/NN-name.mp4   (1080×1920, 30 fps, music, big captions)
        │
        ▼  you: review, (optional) swap in a trending sound inside the app, turn on the promotion disclosure, post

Keep the assets the agent may use in one folder and list them in a short doc it reads first (for example marketing/video/RENDER.md): the named scenes of your screen recording (start second, end second, caption), the stock clips with what they show, and the music tracks with their license links.

Spec format (example)¶

{
  "music": "stock/upbeat-happy.mp3",
  "volume": 0.6,
  "beats": [
    { "kind": "stock", "file": "stock/frustrated-laptop.mp4", "from": 1.5, "sec": 2.0, "text": "Your Downloads folder right now" },
    { "kind": "app", "scene": 1, "text": "Duplicates, side by side", "sub": "Keeps the original, marks the copies" },
    { "kind": "app", "scene": 4, "text": "Everything goes to the Trash", "countdown": 0 },
    { "kind": "split", "sec": 6, "top": { "file": "stock/frustrated-laptop.mp4" }, "bottom": { "scene": 2 },
      "topLabel": "By hand", "bottomLabel": "With the app", "timers": [95, 4] },
    { "kind": "card", "text": "Free on the Mac App Store", "sec": 2 }
  ],
  "outro": true
}

`kind`	Shows	Required fields
`stock`	a licensed stock clip, full screen, caption at the bottom	`file`, `sec` (+ `from`, `text`, `sub`)
`app`	a named scene of your real screen recording in a device frame, caption on top	`scene` (+ `text`, `sub`, `countdown` seconds)
`card`	big text on your brand color	`text`, `sec`
`split`	two sources stacked, labels, optional timers (real measured seconds only)	`sec`, `top`, `bottom`
`slide`	one text slide on a color or stock still (carousel)	`text`, `sec`

Durations come from the beats, so the tool computes the video length; there is no separate timeline to keep in sync.

Rules the agent must follow¶

Only scenes that exist in your recording; never animate a feature the product doesn't have.
Timers, counters and "X GB freed" must be real measurements, or left out.
The first beat already shows the problem or the result (no logo intro).
Captions readable on a phone with the sound off: 60–90 px on a 1080-wide frame, max ~8 words per line.
Music only from the licensed folder; trending sounds are added later inside the app.

What stays manual¶

Posting (each platform's own app or an approved API / scheduler), choosing a trending sound, Stitch / Duet, comment replies, and any format that needs your face.