Hey Charlie, is ThemecoAI still coming this year or early next year? It seems to have quietly been postponed?
1 Like
I can imagine this is quite a task to accomplish. It actually became more realistic now with the release of Gemini 3.0 and Opus 4.5, from what I can tell. Older models need significantly more training for consistent results.
The massive thing here is the fact that there is a system in Cornerstone to handle the “diff” rather than the full state (When Dev Tools is active). So anything that is default is not part of the JSON. As soon as we change a state of a setting, it becomes part of the JSON export. This is super token-efficient because something that has over 400 options suddenly has only a few.
AI has to be trained on all Layout options and states. Right now, we can relatively easily train Gemini 3.0 to accept screenshots and create nice layouts based on them, using Structural elements plus basic ones such as image, heading, button and text. I got pretty nice layouts out of it, as a matter of fact.
The complexity starts when we try to introduce other elements. Most of them have many shared options and states, but each has its own specific “setup” options that need to be progressively disclosed to the LLM, when it decides to use them.
I was away last year so I wasn’t able to accept invitation to participate in AI testing. I can’t wait to see how will you guys solve this. I expect nothing but perfection as always. 
2 Likes
We have some serious progress on it. It’s capable of building layouts and editing any element on the page. It has both impressed me and disappointed me at times. We were originally thinking this would be an end of year release, but we’re having some issues with consistency in output and would rather release it when we’re more satisfied with it. Currently are short term goals are to button up Pro 6.8 and release Cornerstone Forms for early next year. Which would also allow us to train the AI on forms too.
The output already factors in the default values of Cornerstone elements. To limit the amount of data sent and received. There is a little known feature when the dev tools is enabled called “Copy Minimal” when you right click on an element. We use this internally to train.
You’re right that training the AI on each element is complex. We do have most of that done, the problem we are at is putting it all together to build nice looking layouts. We have most templates from Design Cloud in our knowledge base, but it could use a lot more examples. It does astound me that you can say “Recreate the Youtube watch page” or “Change the colors to look like Mountain Dew” and it’ll give you something very similar. I’m sure I’ll only be more impressed as these AI models progress.
Hope this answers your questions. Have a great weekend.
3 Likes
Thanks, that all sounds great. You have a great weekend as well guys.
That’s amazing to hear!
Yeah, consistency. Users are accustomed to re-prompt when vibe coding, but if it’s part of a commercial product, they might be less forgiving.
My method of training is asking it to produce a certain section such as a pricing table.
Once it does, I fix it visually and use Copy minimal to give it back. It then examines it and documents what it has learned.
I did it just for fun, not actual commercial use. But with latest Gemini and Claude Opus, I might get back to it.