A scary premise becomes a finished Short, two taps away
A story prompt becomes a two-path horror Short: Claude writes it, FLUX and MiniMax generate the stills and voice, ffmpeg cuts the VHS edit and you approve it twice in Telegram.
~60–75 sVideo length
Hook, choice, both outcomes
9Images per video
≈$0.4–0.5Cost per video
Images ~$0.25, voice ~$0.1, story ~$0.05
1 dayBuild time
12
Story approval: hook, three setups, the choice and both outcomes.
1Hook: the premise in one line
2Choice: split-screen "1 or 2" decides the path
Where AI does the work
01 / STORY
Story engine
Claude writes each story as strict JSON: a hook, three setup beats, a split-screen choice and two outcomes. Every new premise is checked against a running store of premises already used, so the pipeline never quietly repeats itself.
02 / VISUALS
Stills and voice
fal.ai's FLUX model renders nine images per video with a fixed style suffix so every set reads as one consistent look, and fal.ai's MiniMax model voices the narration in a single deep voice over a shared music track.
03 / ASSEMBLY
VHS assembly
ffmpeg pans and zooms across the stills with a Ken Burns effect, burns in subtitles timed to the narration and layers a VHS treatment on top: film noise, an RGB shift, scanlines and a REC indicator.
Built with Claude Code in 1 day
Before
Every horror short meant writing, images, voice and editing by hand.
After
A prompt becomes a finished video after two taps of approval in Telegram.
[→]/More screens
The finished video card, ready to post after two Telegram approvals.
Under the hood
01Context
Every interactive horror short took a full manual pipeline
A 'which path survives' horror short needs a two-branch story, a consistent set of images to carry it, a narrated voiceover, captions timed to the words and a VHS-styled edit before it is ready to post. Doing all of that by hand for every new short does not scale past a handful of videos, and repeating a premise or a visual set that was already used is easy to miss without a record of it.
02How it works
Claude writes the story, FLUX and MiniMax carry it
Claude writes each story as strict JSON, a hook, three setup beats, a split-screen choice and two outcomes, checked against a running store of premises already used so the pipeline never quietly repeats itself. fal.ai's FLUX model renders nine images per video with a fixed style suffix so the set reads as one consistent look, and fal.ai's MiniMax model voices the narration in a single deep voice over a shared music track.
ffmpeg then pans and zooms across the nine stills with a Ken Burns effect, burns in subtitles timed to the narration and layers a VHS treatment on top: film noise, an RGB channel shift, a vignette, scanlines and a REC indicator.
03Constraints
Two Telegram approvals, publishing stays manual
An aiogram Telegram bot answers only one configured admin chat and drives a two-step approval: it sends the generated story first for a go or a retry, then the finished mp4 for a second approval. Publishing to YouTube itself stays a manual step, since automatic upload needs the account to pass a YouTube Data API audit that has not happened. Cost runs roughly $0.4 to $0.5 per video, split between the images, the voice and the story itself.
ANM SOLUTIONS / CONTACT US
Want something like this?
First 30-minute call is free. We scope it, you decide.
