Recent Posts
Gemini+Figma for UI
abXQRY
No.2660
You can ask gemini to directly access and code your figma designs for free, with a figma MCP server.
>Get a gemini api token(free atleast for the flash model)
>Get your figma token
>Install the gemini cli with npm
>Install and run any figma mcp server, I am using this one
https://github.com/GLips/Figma-Context-MCP
>Connect the MCP server in gemini cli config
something like
[code]
{
"mcpServers": {
"figma-local": {
"httpUrl": "http://localhost:3001/mcp"
}
}
}
[\code]
Now you can run gemini mcp list to check if it is connected.
Now just cd into your project, ask gemini to create a UI with file id(of your figma design) with for example react and add --dir in the end and it will make it for you
abXQRY
No.2661
I have only tried this with one design but it seems to work pretty well. Images don't work in the free tier of figma but that's fine.
I asked it to write it in react inside my nextJS project and it was pretty good, only had to make few changes
ijLHu6
No.2662
>>2660(OP)
Thanks but llm design is always slop
abXQRY
No.2663
ijLHu6
No.2664
>>2663
Shit I read it in reverse i thought you were using llm to design figma ui
abXQRY
No.2758
I have changed my mind regarding this after trying it for some time, LLMs are shit
Write code by hand
Sc9mrF
No.2759
>>2758
kek
vc5AK+
No.2760
>>2758
e+DKlo
No.2761
>>2758
Lmao, this deserves a space in /g/ banner
HqpwXT
No.2762
>>2758
LLM generated code is often way too optimal and concise and use latest syntax of whatever language you are using, which even sometimes confuses me what that code even means.
abXQRY
No.2767
>>2759
>>2761
>>2760
>>2762
It isn't simply worth it, they keep making simple mistakes again and again that you have to find and fix, I was using gemini 2.5-pro which was one of the best till like 2 weeks ago and it was still so trash. Gpt 5.1 was slightly better and gemini 3 pro is even better but they still make basic mistakes which doesn't make sense when thousands of people are probably promoting them for similar things
abXQRY
No.2768
*prompting.
Ig they optimize these models for questions like "build me a x app using y stack" but as soon as you get any specific like asking it to write the code using a specific scheme and more precise instructions, it can't follow them
