My automated subtitle system, and the 59.9% problem it hasn't fixed
Most viewers watch shorts with subtitles on. Mine are generated, styled, and placed automatically. This post covers how, and then the honest part: the one retention number that all this automation has not moved.
What the subtitle system does
For every video, the pipeline:
- Splits the script by scene and times each line against the measured duration of the narration audio, not an estimate. Millisecond drift between voice and text is what makes a video feel broken.
- Wraps lines by word units. A subtitle that cuts a word in half reads as amateur in any language. Getting Korean word-wrapping right took longer than building the entire background music system.
- Applies one of five preset styles (clean, bold, neon, variety-show, emotional) with size and position presets, so a channel keeps one consistent look without per-video decisions.
Zero subtitle work per video. Across 60+ shorts, that is hours of timeline fiddling that simply never happened.
The number automation didn’t fix
Here is my least favorite receipt:
| Metric | Number | Source |
|---|---|---|
| First-second swipe-away rate | 59.9% | YouTube Studio |
| Status after months of iteration | still stuck near 60% | YouTube Studio |
Six in ten people leave in the first second. Subtitle quality, sync precision, styling: none of it matters in second zero, because the viewer hasn’t read anything yet. What decides that second is the first frame and the hook, and that is a creative problem, not a pipeline problem.
I’m saying this plainly because automation content usually implies the opposite. Automation removed my production bottleneck. It did not make the videos better at stopping thumbs. Those are different jobs.
What you can use today
- Check your own first-second drop in YouTube Studio (retention graph, the very first cliff). Most creators have never looked at it.
- If it’s high like mine, spend your next hour on first frames, not on editing polish. The polish is invisible to someone who already left.
- Whatever subtitle tool you use, sync text to measured audio timing. If your tool estimates, you will drift.
Next receipt in this series: the hook experiment I ran across 40+ shorts to attack that 59.9%, and which title styles actually survived it.
Here are the receipts.