The stack at a glance
| Job | Tool | What it produces |
|---|---|---|
| Research and script | Claude | A sourced script, scene list and voice-over text |
| Voice | ElevenLabs | One narrator take, with sound effects written into the script |
| Scenes | GPT Image (in Figma Weave) | About 13 drawn stills per Short |
| Motion | Kling 3.0 Turbo (in Figma Weave) | 3 clips: the hook, the twist and the best gag |
| Music | Lyria 3 Pro (in Figma Weave) | An original track timed to the voice |
| Timing | Whisper | Word-level timestamps for captions |
| Edit | Remotion | The final 9:16 video with captions |
A typical episode uses about 230 Figma Weave credits for images, clips and music. All in, $1 to $3 per Short, plus subscriptions.
Script: Claude, checked by people
Claude writes from a topic library and a fixed format: a hook in the first 2 seconds, 11 to 14 scenes, a twist, and a teaser for the next episode. Every fact needs two sources (Etymonline, Wiktionary, the OED). AI is fast at drafts and bad at etymology, so the fact check is not optional.
Voice: ElevenLabs
One narrator voice tells every story; the characters never speak. Sound effects like screams or crunches go into the script as ElevenLabs tags, so they land in the same take. Recording and approving that take is the one step a person does by hand.
Scenes: GPT Image through Figma Weave
Each scene is a still image, generated from three things: a locked style paragraph, reference sheets of the characters in the scene, and the scene's action. We run it through Figma Weave so every image, clip and music track for an episode lives on one board.
Two rules that saved us the most rerolls:
- No text in generated images. Every word on screen is added in the edit, so spelling is always right.
- Never more than 4 characters in one image. Crowded scenes get split in two.
Motion: Kling, only where it counts
Full AI video for every scene would cost more and drift off-model. So each Short gets just 3 animated clips from Kling, each starting from its scene still: the hook, the twist and the funniest moment. The rest are stills with gentle camera moves in the edit. Viewers rarely notice, and the characters stay consistent.
For bigger ad work, like our Sitechecker ad, we use Kling 4.0 with saved character Elements for every shot.
Music: Lyria 3
An original track per episode, generated to the voice-over length with a small sting at the twist. Original music means no copyright claims, which matters for monetization.
Edit: Remotion
Remotion builds the video from code: it reads the scene timings and caption file, places each still and clip, adds word-by-word captions with the current word highlighted, ducks the music under the voice and exports a 1080 x 1920 file. The last frame matches the first so the Short loops.
Because the edit is code, a new episode is mostly new assets, not a new edit.
What we would tell you before you copy this
- Setup is the real cost. The style lock, cast sheets and script format took longer than the first five episodes.
- Keep your own reference sheets. AI tools change every few months. With your own cast and style files, you can switch models without the show changing face.
- Tools are not the moat. Anyone can buy them. Topics, research and a recognisable cast are what make a channel.
The full story of setting up the channel is in our faceless channel guide.