Lol, I have been exploring in the opposite direction. Given a set of basic CRUD endpoints for various data structures, including links to other structures. Generate a basic UI for them, with guidance on added ui to generate based on input and data oriented interaction specs.
This is cool though, I think there are going to be an explosion of different higher level development patterns that will be useful in a lot of different situations that AI will enable. It will be interesting to see which ones get wider adoption, and which stay useful but in smaller focused niches.
Domain driven design or data driven design has it's benefits for sure.
A scenario that comes up is the combination of understanding the problem space, use of tools and what the developer knows/needs to learn can shine their path in different direction to get to the outcome.
At my org we have been using the Figma MCP server to generate code from Figma designs. Spent about a day writing rules to keep the AI on guardrails but it's been very smooth sailing since then and any generated code needs minimal changes.
As far as rules, it's very dependent on the codebase you are working in. My recommendation for rolling something like this out is take about a day or so and try to build UI components with it, and then add to your CLAUDE.md/.clinerules/etc as the AI screws up, and there will be a ton of screw ups. Funnily enough the first rule I had to write was to tell Cline to use the MCP plugin because it kept trying to run a fresh copy and could never connect.
I guess it's good for "Dreamweaver 2.0" in the sense that it produces plain HTML. But my understanding is most serious web applications use a framework of some sort, such as React, Vue, or Svelte, and each is opinionated.
It's so much harder to make very accurate image to plain HTML conversion than plain HTML to react et at. LLM are very good with translating between different web frameworks.
Oh! What about CSS frameworks like Tailwind, Bootstrap, etc.?
How on earth was it able to generate and clone Images just from UI? Like did it get the source of the URL as well, where it could link those images or what did it actually do there?
Readit News
This is cool though, I think there are going to be an explosion of different higher level development patterns that will be useful in a lot of different situations that AI will enable. It will be interesting to see which ones get wider adoption, and which stay useful but in smaller focused niches.
https://rjsf-team.github.io/react-jsonschema-form/docs/
A scenario that comes up is the combination of understanding the problem space, use of tools and what the developer knows/needs to learn can shine their path in different direction to get to the outcome.
As far as rules, it's very dependent on the codebase you are working in. My recommendation for rolling something like this out is take about a day or so and try to build UI components with it, and then add to your CLAUDE.md/.clinerules/etc as the AI screws up, and there will be a ton of screw ups. Funnily enough the first rule I had to write was to tell Cline to use the MCP plugin because it kept trying to run a fresh copy and could never connect.
So many things wrong with this sentiment.
Define “serious.”
Do people think it’s just not possible to make web apps without frameworks?
Take the output from this tool and adapt it to whatever framework you want … if you feel that’s necessary.
How on earth was it able to generate and clone Images just from UI? Like did it get the source of the URL as well, where it could link those images or what did it actually do there?