How We Automated a Daily AI Video (And Everything That Broke First)

For several weeks now, a routine has been running entirely on its own: every morning, a roughly 15-second vertical video gets generated, edited, and delivered to my inbox, without me opening a single editing tool. Internally we call it "Remy Daily Video." It's not a template you fill in. It's an agent that takes a day's AI news story, turns it into a script, stages it, renders it as video, and delivers it.

What I didn't see coming was how bad the first version actually was. This article tells both halves of the story: what's running today, and everything we had to break to get there.

What the routine actually does

The format is built for Reels/Shorts/TikTok: 1080×1920, 30fps, a dark aesthetic with a gold accent, an animated stat counter on open, film grain, a particle field in the background, and a progress bar at the top. Each video covers one single story, not a digest.

It's not a template you refill every day. It's an agent that writes, stages and delivers, untouched.

What's under the hood

Three pieces stitched together, no magic:

The real failures before it worked

This is the part nobody tells you in "I automated my content" posts. The first version technically worked, it rendered a video every day, but it was bad. Nobody watched it to the end.

The retention crash

A retention diagnostic hit us hard: 75% of viewers dropped off within the first 3 seconds, and only 23% watched to the 29-second mark. The video was too long, too slow to start, and the message got diluted across too many scenes.

The fix was drastic: total runtime cut to 12-15 seconds, a rigid 5-scene structure locked in, hook, three stat beats, call to action, and one transitional scene removed entirely because it carried zero data. It "looked nice" but did nothing for retention.

The hook that kept missing, over and over

The opening scene, the one that decides whether someone stays or scrolls, had to be fixed multiple times in a row, on details that look minor in isolation but kill a hook when they stack up:

• The wrong subject named. One version credited Anthropic as the protagonist when Google was actually the story, an opener pointing at the wrong actor instantly loses credibility.
• An unattributed number. Another version opened on an abstract, unattributed statistic instead of directly naming the company or person involved, less punchy, less credible.
• A headline spilling to two lines. A hook that wraps to two lines instead of one loses the gut-punch effect, the rule had to be forced: one line, one message.

The message fragmented across too many stats

The deepest problem wasn't visual, it was narrative. Early versions tried to cram 4 different stats into the 3 middle scenes, a pile of facts with no connecting thread. The fix: one single narrative line per video, structured in three beats, a claim, the paradox that contradicts it, the real read behind both. One point, fully explained, instead of four points barely touched.

The infrastructure bugs, less glamorous but just as real

And then there was the plumbing, invisible but blocking: conflicts when the pipeline tried to re-publish a video that already existed for the same date, and headless browser installs that failed depending on the execution environment. Nothing dramatic, but any one of these bugs left unfixed would have been enough to break the routine every other morning.

What it produces today

The pipeline now runs every day with zero manual intervention: automatic morning trigger, render, publish, notification. Every video covers one single AI story, framed as a paradox to resolve rather than a raw fact, the format that, after the fix, holds attention instead of losing it in 3 seconds.

This is exactly the philosophy we sell at Asymmetriq: not a tool you open and close, but an agent that runs while you sleep and only asks for approval. The difference here is that we built it and broke it ourselves before ever selling it to anyone.

Frequently asked questions