No Swipe

My kid watches football highlights on the sofa every evening.

He doesn't care who made them. Doesn't read the channel name. Doesn't check the description. He watches for about four seconds and decides. If the footage stalls, if it drifts into story mode, if anything feels off — he swipes. No explanation. No second chance.

He's like Caesar deciding a gladiator's fate. Thumbs up or thumbs down. Except faster, and with less deliberation.

I was sitting next to him last week, on my phone, building an automated football video engine with my AI agent. He was watching highlights. I was building a system to make them.

And the thought landed: what if he ends up watching mine?

Not "what if I tell him and he thinks it's cool." Not the dad card. What if the content is good enough that it just shows up in his feed and he watches it. All the way through. Without knowing I made it.

That became the bar.

The World Cup starts in ten days. That's the window I'm building for — three channels, nine social accounts, automated pipeline, daily content. The whole thing runs on n8n workflows and a render service I built with my AI team.

But the World Cup isn't the real reason. The World Cup is the excuse. The real reason is that kid on the sofa.

I rendered the first test video and the green screen keyed out the football pitch. Grass is green. Green screens are green. The AI didn't see the problem.

Second render: switched to magenta. The dark elements bled into it. Captions got a visible tint. Unwatchable.

Third: fixed the compositing by ditching it entirely. FFmpeg text overlays instead of fancy templates. Worked. But the captions stacked on top of the title card. Two text layers at once. Visual noise.

Fourth: suppressed captions during the title. Better. But the clips were cut at fixed five-second intervals. A goal celebration got sliced in half. A number card from the broadcast made it into a highlight clip.

Fifth: switched to scene detection. Let FFmpeg find the natural camera cuts. The clips made sense now. But the video ran out of footage before the voiceover ended. Two seconds of black screen at the end.

Sixth: trimmed the voiceover to finish before the footage. Added crowd noise from the actual broadcast audio instead of a generic stock track. Matched the channel badge to every frame.

Six renders. Six failures. Each one something the AI built perfectly according to spec — and each one something my kid would have swiped past in under a second.

The system can script, voice, time, render, and publish a video without me touching it. It can do this twenty times a day. It's fast and it doesn't get tired.

But it can't watch the result and feel that something's off.

It can't sit on a sofa and know — without words, without analysis — that a cut feels wrong. That the energy dropped. That the transition between two clips broke the momentum. My kid can't explain it either. He just swipes.

Taste is the last human job. Not because AI isn't good enough. Because taste isn't a skill. It's a feeling. And the feedback loop for feelings is a seven-year-old with a phone and zero patience.

He doesn't know I'm building this. He might never know. Maybe someday he reads this blog and pieces it together. Maybe not.

The point isn't the reveal. The point is that the content earns its place on its own. No "dad made this" card. No charity views. If he watches someone else's highlights every night, why can't he watch mine?

That's when I'll know I made something good. Not when the view count moves. Not when the algorithm picks it up.

When my kid watches a video all the way through and doesn't swipe.