01
It sees the video.
Every 2.5 seconds a frame is sent to the vision model. Face bounding box, shot type, gesture, emotion — all fed back into the crop so the speaker never drifts out of frame.
No.001 · Made in Cyprus · Updated today
Long video in.
Watchable clips out.
Not another OpusClip. This one sees the video — gemma4-31b reads every frame, whisper-large-v3 catches every word, the crop tracks the speaker's face across every shot. Six-axis score, per-platform virality read, refuses to ship a mediocre clip.
6
scoring axes. Hook, emotion, clarity, payoff, retention, virality. Every clip earns its slot — nothing ships on a whim.
03.
How it thinks
Three ideas set this apart from every OpusClip-shaped tool.
02
It scores by audience.
Six intrinsic axes plus a platform read — TikTok, Shorts, Reels get different scores because they behave differently. The pipeline tells you where the clip actually belongs.
03
It refuses mediocre clips.
Reject-first prompt. Rather ship one strong clip than five mid ones. Intros, filler, sponsor reads, throat-clearing — all trimmed automatically before you see the result.
“
Rather return 1 great clip than 5 mid ones.
— analyst system prompt
04.
The stack
- 01Ingestyt-dlp · 1080p mp4 + isolated audio track
- 02Transcribewhisper-large-v3 · segment timestamps
- 03Analysegemma4-31b · reject-first prompt · 6 axes
- 04Seegemma4-31b vision · face & shot per 2.5s frame
- 05DirectLLM plans camera moves per clip
- 06Renderffmpeg 9:16 · head-room-aware crop
- 07SubtitleASS burn-in · max 3 lines · keyword accent
- 08Verifyvision QA · 0–100 composition score