AI agent for reels
An agentic pipeline for short videos: topic, facts, script, voice-over and editing with motion graphics
Task
A production pipeline for a short-video channel. An agent picks a topic and gathers a fact dossier, a separate call writes the script, and the voice reads the whole video in a single request. Editing is done by a second agent: it writes the plan and the animations itself, watches the rendered frames and fixes them. The finished video goes out for upload to the platforms.
How it works
- Topic and fact dossier — the release agent
- Script — the prompt that won a blind comparison
- Voice — ElevenLabs, word timings by speech recognition
- Editing — the agent writes the plan and animations in Remotion
- Checks: pace of cuts, share of footage, subtitles in two lines
- Upload to the platforms
What had to be solved
The voice sounded in fragments
The script was voiced in pieces, and the intonation broke at the joins — the video sounded like a splice. Now the whole script goes to voice-over in one request, and speech recognition against the known text restores where each word falls. It also provides the timings for subtitles and editing.
Commits
- 713da6a29.09.2026the release is edited by the arena winner, with the same agent as in the arena
- c1b92ac25.09.2026voice in a single request — separate pieces sounded fragmented
- 6813cf825.09.2026the arena checks the choice of facts — the model was dragging everything into the video
- 3a3856629.09.2026the subtitle plate is never taller than two lines
- e3543d428.09.2026jokes are the video’s style, the voice-over doesn’t waste characters