I Made My Website Promo With AI Agents

What if you could direct a major promo or product video through one ongoing session with AI agents? With the workflow I built, I pretty much did, in my spare time while I was also working on other projects. If I had put this together through the traditional production process I'm used to, I would have expected it to take weeks.

I want to show you how it came together, because I think this says a LOT about where creative work is going.

Two things made this possible: Claude Opus 5.5 and the multi-agent architecture I built with my internal Hermes / OC agents.

Here's how we made it:

1. I directed it. I chose the work, the story, the visual direction, the claims, and the moments that needed more time on screen. I kept reviewing cuts and saying what felt right or wrong. That part took real judgment. AI made it possible for me to act on those decisions much faster.
2. The agents did the production work. My other agents helped gather and organize the media. I worked with Claude in Cowork to build the titles, graphics, animation, transitions, and edit as code. The picture was designed like a web page to match my site's type and colors, then HyperFrames rendered it frame by frame as a 1080p video. A full render took about four and a half minutes, so I could look at a version, give notes, and see another one without waiting for a whole production chain to turn over.
3. ElevenLabs helped make the music. We didn't get there minute one but we got close. The devil is in the details though and I wanted to nail it. Claude generated six scores in code, then three more directions using recorded samples. I kept hearing something too synthetic. I eventually sent a picture-only cut into ElevenLabs and used its video-to-music tool. The first pass was clean, but it felt too safe for the film.

The edit gave us a way to be much more specific. The opening cuts happen every 0.4 seconds. That is one beat at 150 BPM, so we asked ElevenLabs for music at that tempo with the downbeat at the very start. The next result had the energy I wanted, but the track still had to be placed against the picture. Claude measured where the hits were, adjusted the edit, and even changed a few picture cuts by fractions of a second so they worked with the music. I listened and made the call on whether it actually felt good. A timing calculation can tell you a cut is on the beat. It can't tell you that a one-beat stop feels like an accident.

We caught other things along the way: the wrong image on the first frame, a transition baked into one of the opening shots, and a “Draft” label that absolutely did not belong in a finished video. We made eight versions because I could ask for a change in a chat, but I still had to watch what came back and catch the stuff the system missed.

I had already built a multi-agent system that could gather and hand off assets while I kept directing the cut with Claude. A LOT of production time can disappear into finding files, checking versions, and getting the right material into the edit. My agents could handle that while Claude built graphics, rendered cuts, and ran technical checks. I could keep working on other projects, then come back, review the work, and move the next version forward.

For a self-produced website film, this level of access is enormous. I could experiment with motion design, editing, music, and finishing without booking a separate team for every pass. I still value the people who do those jobs professionally, and there are projects where I'd want them. Here, I could direct work across all those areas without becoming the person operating every tool in the pipeline.

I still had to listen to the music, make calls about story and credits, and check every claim and clip before putting this on the website. This is a very powerful way to work, and it makes the producer's job of knowing what to look for MORE important.

I want to keep testing how far this can go. The next cut will be even better.

Stephen