I Spent 10 Days Teaching an AI to Make Launch Videos That Aren't Embarrassing

I'm a solo founder. I can build a product. What I can't do — what I never have time for — is the marketing: the launch video, the GIFs, the posts, the articles a good product needs and quietly dies without. So I built a tool to do it for me. It's called FoxPlug, and for about ten days straight it made videos I'd have been embarrassed to send anyone.

This is the honest story of those ten days, because the mistakes are more useful than the wins.

The idea was simple. The output was garbage.

The pitch is easy to say: paste a URL, get a launch video. Under the hood, FoxPlug pulled a product's public images and stitched them into a 30-second film with captions and music. Clean concept. The first videos were a paragraph of marketing copy laid over a stock gradient, set to a snappy beat. Walls of words. No product. If you'd shown me one and asked "want that for your launch?", I'd have said no.

The first trap was polishing the wrong thing. I kept making the text nicer — fonts, captions, transitions — when the real problem was that there was no product on screen at all.

"It doesn't show the builder actually working"

The turning point wasn't my idea. I gift these videos to founders on Product Hunt who launched without a demo, and I went back and asked, point blank: what would make this good enough that you'd actually put it in your launch gallery?

One reply did more for me than a week of my own guessing:

"The video looks nice, but it lacks actual usage of the builder — the drag and drop, moving and filling inventory, the exports. That's the whole point of it."

That's the whole thing, isn't it. A launch video isn't a pretty title card. It's the click, the drag, the "oh, that's what it does." I was making postcards when people wanted to see the product move.

The fix that mattered: use the real product, or don't pretend

So I rebuilt the engine around one rule: use the founder's real product screenshots, and refuse to fall back to generic animation when real ones exist. Sounds obvious. It was five bugs deep.

The images makers give you aren't clean screenshots — they're wrapped in marketing frames (a big header, annotation callouts). My engine rejected those wholesale and defaulted to stock motion. Meanwhile the GIF side of the same product cropped them to the app window and used them perfectly. Same input, opposite result, because one cropped and one rejected. Once I made the video reuse what the GIF already figured out, the tours finally showed the actual product — a real database, a real dashboard — not a lock icon on a gradient.

I also added a gate that blocks a weak render before it publishes: blank frames, the same asset used twice, too much text, a cursor that never moves. Fail, and it doesn't ship. And every render is stamped with the exact code that made it, so I can prove what produced what instead of guessing.

The bigger lesson: a recorder is a commodity; what you do with the recording isn't

Even with real screenshots, a video built from static images has a ceiling — it can't show the drag-and-drop. So the next move was obvious and a little humbling: let people record their live product — a 30-second click-through, in the browser, no software, no mic — and turn that into the film.

But I had to be honest with myself: a screen recorder is a commodity. Loom does it. macOS does it free. Nobody needs another one. The reason to record inside my tool can't be the recording — it has to be everything after. One real recording → a launch video and a GIF series and a sticker set and stills and a blog post and native posts for every channel. One input, a launch's worth of content out. That part isn't a commodity.

What I got honest about (the part most build-in-public posts skip)

Ten days in, here's my traction: a handful of signups, no revenue, and — I checked — time-on-site was dropping on the days I thought the videos were getting better. A third-party "usefulness" audit scored the project 37/100: genuine pain point, modern build, two big zeros — reach and traction.

That stung, and it was the most useful data I got. It told me the product isn't the problem; distribution is. Every early signup came from one channel: gifting real, white-label launch videos to founders who didn't have one, on the day they launched. Slow. Hand-earned. But it's the only thing that's converted anyone — and it works because I do something useful for someone before asking for anything.

Where this leaves me

I won't tell you it's fixed. (I've learned not to say "fixed" — I've said it too many times.) The videos are real now. The recorder works. The repurposing exists. Whether "record your real product, get a whole content kit" is the thing that finally moves those two zeros — that's the next ten days.

If you're building in public and staring at the same wall — you can ship, but you can't find the hours to tell anyone — that's exactly who I'm building this for. Because it's me.

Building FoxPlug in public