Produce Instagram Reels from a script with generated images, voice and subtitles

The small team behind the Korean app Harudle built a Codex-driven Reels pipeline that cut their average production time per Reel by about half and raised their own quality score by about 30%.

Done withCodexinElevenLabs· Oct 24, 2026

What they did

The Harudle team, students building a four-panel comic app, each had to post one Instagram Reel a week to promote it. Making one by hand in Vrew took about two hours. They turned their existing comic-making process into a Codex skill that generates each scene as its own image, including a hook image for the first three seconds and a call-to-action image at the end, with shifting camera angles to keep viewers watching. Narration comes from a paid ElevenLabs voice after open-source TTS sounded too robotic. stable-ts with a faster-whisper model finds where each word of the script lands in the audio, and FFmpeg lays the images and timed subtitles over the voice track as a vertical MP4. The cover is made with a computer-use step. A teammate reviews and approves each draft before it is posted.

How it went

By the team's internal scoring, average production time fell by about 50% and the average quality score rose by about 30%. They say the main gain was getting their time back for development: they plan a Reel briefly and leave generation to the pipeline. Posting is still manual. Views went up, with earlier Reels reaching 50,000 and 350,000, but follows and sign-ups from those views stayed low.

Worth knowing

The post is in Korean. Generating each scene separately means the drawing style can drift between frames; the team accepted that for more control over each shot. The voice needed a paid TTS plan to get natural stress and intonation. Getting more views did not turn into more users without a stronger call to action.

Try it

With Codex· 5 parts to fill in

Build me a short-video pipeline for [my product]. Given a [30-second] script, generate one image per scene in [my visual style], with a strong hook image for the first three seconds and a call-to-action image at the end. Create the narration with [my TTS service], align each word of the script to the audio for subtitles, and assemble a vertical MP4 with FFmpeg that shows each image and its subtitles at the right time. Also make a cover image. Save everything to [output folder] and stop there: show me the draft video and cover for approval before anything is posted.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related