Most content fails before it is made. The topic was crowded, or the opening lost people in the first few seconds.
This engine moves the evidence to the front. It is three Claude Code skills that run in sequence. The last one, viral-hook, is open source.
- viral-spotter finds outliers. It scores each video as its views divided by the median views of the channel. 3x is an outlier, 5x is strong, 10x is gold. Shorts and long-form are scored apart. Then it takes the winner apart: the hook, the topic and the format.
- topic-radar ranks the backlog. It combines outlier signals, live news and a YouTube demand check. Weights: outlier ×2, timeliness ×2, contrarian angle ×1.5, channel fit ×1.5, plus monetization. A saturation penalty drops topics that AI channels already flood.
- viral-hook scores the opening. It tests the first 3 seconds against five hook types and a six-point gate. It returns a winner and two alternates. It draws on an OpusClip study of 34,635 clips: the algorithm decides in about 1.5 seconds, so the hook must land by 3.
After that, two more checks run: one on retention across the whole video, and one on whether the title and thumbnail ask one connected question.
Source on GitHub (viral-hook): github.com/abhilashme/viral-hook