A lot of the agentic coding I do these days is client work, and most of it I can’t show. It’s commercially sensitive. So when something small and harmless comes along that shows how I actually think about this stuff, I’ll take it. This one involves a mini trampoline.

On Saturday morning I was handed a job: set up an Everfit 48 inch hexagon rebounder. Should be simple. Frame, legs, mat, a bag of bungees. Then I got to step 6 of the user manual, fitting the bungees, and stopped.

Here’s the step in full:

  1. Use the assemble tool to fix the bungees as below,
  1. Place one end of the bungee on the metal loop, insert the assemble tool into another end of the bungee, and make it go through the frame, push the assemble tool forward to make the bungee go inside of the metal loop too. Pay attention, when using the tool, the big end of the tool should face the direction of the metal loop.
  2. Assemble one of the bungees from each side of the frame to ensure that the bungees on each side with even force.
  3. And then assemble the rest bungees one by one, as shown above.

That’s it. Three sentences and one small grainy photo of a hand holding a wire tool. And there’s no metal loop. Anywhere. The mat has fabric tabs around the edge, the bungees are elastic rings with a plastic ball on one end, and there’s a single wire triangle that I assume is the “assemble tool”. No hooks either.

The YouTube detour I didn’t take

If you’ve set one of these up before, you’re probably wondering what the fuss is about. Fair enough. It wasn’t obvious to me.

These days YouTube is the go-to for this kind of thing. There’s a how-to video for everything. The trouble is there’s a wide variation of rebounders that look almost identical but attach the bungees differently. Hooks, springs, different tools. I gave a few of them a few seconds each, but none looked like what was on the carpet in front of me, and I didn’t have the patience to sit through them in detail hoping one would turn out to be close enough.

I could have kept searching, but I didn’t want to do that work. I just wanted to understand what was needed and get the job done. So I decided to experiment.

Building an app instead

I took a photo of the manual page, gave it to Claude Opus, and dictated what I wanted. Research this exact model. Don’t confuse it with other trampoline types. Build me an HTML page that explains how to fit the bungees far more clearly than the manual does, step by step, with a 3D diagram for every step using three.js. Then I told it to get started while I went and took photos of the parts.

This is the part I find interesting. When you do agentic coding every day, it changes what you reach for. It’s a way of thinking, and that way of thinking leads me to do crazy things like build an app to explain how to attach a bungee to a rebounder.

Over the next few minutes I sent a few photos of the parts: the wire tool, a bungee, one of the fabric tabs on the mat. Claude found the manual for a sister 50 inch model with eight photos of the same bungee and tab. It also found the same wire triangle sold separately. From that and my photos it worked out how the whole thing goes together, which the manual never really explains. It got there from a poor scan of a manual and some phone photos of loose parts on carpet. That level of visual perception still surprises me.

When I next checked in, about 25 minutes after my first message, there was a nine-step interactive guide sitting there. Each step has an animated 3D scene you can drag around and zoom, with the instructions beside it and which hand does what. There’s a diagram of the order to work around the frame, a troubleshooting section, and a straight answer on what it was sure of and what was its best reading.

Claude publishes these as artifacts, so I opened it on my phone and propped it up next to the rebounder. The bungees went on one after the other, and the thing was finished.

That app had a lifespan of one Saturday morning. It was built for one person and one problem, and used once. That’s fine. Software is now cheap enough to build for a single moment.

Then I made a video about it

This was the second experiment.

I already use three.js with agents, so the guide wasn’t new territory. What I’d never done was use HyperFrames, an open-source HTML-to-video framework, or have ElevenLabs narration generated automatically as part of an agentic build. So once the rebounder was done, I asked for a video.

I dictated the brief, same as the first time. Introduce the problem, show the manual excerpt and my photos, then demo the guide end to end with the 3D animations. Add subtitles and narration in an Australian male voice, close to me but not me. Use my intent, not my rambling wording.

While it was building, I changed my mind twice about what the video was really about. I asked it to introduce me and finish with a soft ask. I also asked for visible mouse drags in the demo, so people can see the 3D is live and not a recording. For the narration, I gave it an ElevenLabs API key limited to the endpoints it needed and set to expire after one day.

It wrote a 27-segment script, generated narration with character-level timestamps, and timed 148 captions to the speech. Then it drove the actual guide in a headless browser, frame by frame, so the demo lines up with the voice. The final cut is 6 minutes 24 seconds at 1080p.

It took longer than the guide did, but it all ran asynchronously and I got on with other stuff. I didn’t care about the details of how it got there. I cared about the result, and then I had a video.

All up, my input across the whole day was a handful of short messages and photos. The video took around three hours end to end, nearly all of it the agent working on its own.

Is it any good?

Good enough. I picked an Australian voice because it’s close to my accent, but I didn’t bother with a voice clone that would actually sound like me. I might do that one day. It wasn’t needed for a quick turnaround thing like this.

There’s an occasional glitch in the audio at the end of a phrase. Some people will look at the whole thing and call it AI slop. That’s fair, but it’s also informative, and given what it’s for, it isn’t worth an iterative loop to polish it. You get to make that call when the first version costs this little.

I’ll use HyperFrames and automated narration again. Like everything in this space, it needs more experimentation before I really know what it can do. You can always tune the result to taste, and it’s still quicker than doing it from scratch by hand. Speed isn’t really the point, though. This kind of capability opens up scenarios we just haven’t considered before.

What this is really about

The trampoline was just the excuse. What I care about is what happens when building a small piece of software takes less effort than finding someone else’s explanation.

I can’t show you most of my work, but this is how I think about it. I’m in my fifth decade of writing software, and these days most of it is agentic development. If you want help building apps for engineering or business, whether they’re disposable, short term, long term or systems of record, get in touch.