Some notes from the last few months, in no particular order.

Nobody wants to choose a model. We gave users a dropdown early on. Almost everyone left it on the default, even when another model was clearly better for the job. So now an agent reads the request and picks. You can still see what it's choosing from on the models page.

Text in images is its own problem. Posters and thumbnails with readable text need a different model than photoreal product shots. Routing on "does this need text" alone fixed a lot of complaints.

Video prompts need more structure than image prompts. Camera movement, duration, what changes between start and end. Users never write that, so the agent fills it in.

Consistency beats quality. People forgive a slightly worse image if the character looks the same as last time. That's why the reference image tool gets more use than anything else we ship.

Show real output. Our showcase converts better than any copy we've written.

If you're building something similar, I'm happy to compare notes in the comments. And if you just want to try it, new accounts get 50 free credits.