Good tools for getting AI agents and chat to recommend **new css features** ?
I feel like AI coding tools are always either suggesting...
1) outdated css advice
2) using javascript where modern css will do
3) importing packages and dependencies when raw css will do
Does a good tool already exist, or
do we need to have a bot watch all Kevin's videos and RAG them?
48 Replies
Honestly, you're not going to find an AI with recent knowledge. That's one of the many downsides of LLMs: they have to retrain on new data, so they're always going to be at least a few months outdated, if not moreso.
The best option to know what's new in CSS land is watching Kevin's videos, following some influential CSS people, and just poking around on MDN to see what's new.
well, you described the trends over time as well
and ai will follow the trend
As far as I know there's no official changelog for CSS to use.
there's the spec
when things leave the draft, should go to the spec itself - the non-draft one
Right, but the spec only updates itself but doesn't have a changelog to tell you what was added
but it doesn't matter much: if 99% of the world used
outdated-technique-xyz, the ai will 99% likely recommend ityes, this is the problem I'm having
that's how ai works
and you will always have that problem
I feel like that's either RAG is for. (maaaaaybe finetuning instead)
I hate to pull a stack overflow, but I'd genuinely recommend just not using Ai. It's slower in the end, it dulls your ability to reason about code, it's expensive even when they're giving it away for free now, you won't learn if you're relying on Ai, it's trained on data that's obtained unethically and it's an unprecedented environmental disaster
Like, not using it at all
I largely agree with most of that. But notably I'd say in ref to your statement "it's slower in the end" I'd say "for now, given the current tooling". Because it does generate slop, with way too much js to do styling. But if new tool come on line that may change. I'm trying to figure out if those tools exist.
The biggest issue with LLMs is what I like to call "historical interia". Things that are popular will have a lot of resources that the LLM consumes. Making the LLM more likely to suggest said thing. Which means more people will be using it thus making LLMs more likely to suggest it.
Popularity breeds more popularity.
yes, I 💯 agree. "historical inertia" is the problem. are there any tools to help overcome them?
And if you have a look around StackOverflow and other sites there's a lot of crap code ("my crap code doesn't work, halp!") and that's the stuff LLMs train on. Or tuturials for beginner projects that only have the bare minimum to work but without the necessary follow-up with production-ready code. That means that LLMs are gonna be a lot better at tutorial-level/prototype code but will crumble when you need to, for example, add authentication or handle PII
I don't think fine tuning a custom model would be of benefit, as too difficult and restrictive.
Real, live, people. Ask questions here. Watch Kevin's videos. Take courses from trusted experts in the field. Read MDN. Start a blog to chronicle your journey. That way you help others who find your stuff
But maybe a RAG system setup via MCP?
People are going to be the best point of reference for anything, especially new stuff
I don't know what RAG or MCP is
Don't get me wrong, using LLMs can be useful if you do it smartly. Most of my LLM interactions are thus:
I want to doThat way I'm not asking for code but for suggestions on ways to implement things. I ask for lists of steps to do<x>and I plan on implementing it via<y>. However, I'm concerned about<z>. What do you think?
<x> so it gives me what I need to code it myself.
Sometimes I'll give it my prototype code and ask for it to evaluate. But I'm experienced enough to know when it tries to lie to me. I know, for example, that you can't for...of a number. You need an iterable. So if it tells me for...of a number I know that's a dumb ide
-# ok, very bad example but I hope you get the ideaRAG=Retrieval Augmented Generation
MCP=Model Context Protocol
to stay updated, wouldn't this basically need a 2nd model just to feed the 1st model?
MCP is the easier of the two to explain. do you use any AI tools?...Copilot, Claude, etc..
GPT is the only one I currently use
chat gpt or codex?
perhaps, but you could argue that's "an implementation detail".
agents are getting where they can be experts in a field and collaborate with one another. Although not popular yet, A2A (Agent2Agent Protocol) will be used for this
GPT, the web interface. Free version
What IDE doe you use?
WebStorm
Di you use it's AI chat?
Nope. I separate out my IDE and LLM
On purpose. If I’m too lazy to cmd + tab to the LLM then it’s not important enough to ask it
That keeps me more honest and more of the code my own
Chat GPT doesn't report mcp servers at the moment. If you want to try AI chat in a jet brains product here is the resource for setting up MCP servers https://www.jetbrains.com/help/ai-assistant/configure-an-mcp-server.html#
AI Assistant Help
Configure an MCP server | AI Assistant
I don’t, as I said already
if you want to try
if you're comfortable with the terminal, gemini cli is probably the easiest one to get up and going https://blog.google/technology/developers/introducing-gemini-cli-open-source-ai-agent/
Google
Gemini CLI: your open-source AI agent
Free and open source, Gemini CLI brings Gemini directly into developers’ terminals — with unmatched access for individuals.
Probably the first MCP server I'd start off with is context7. Which is trying to reduce hallucinations by providing llms with up to date information on Frameworks. https://www.youtube.com/watch?v=323l56VqJQw
AI LABS
YouTube
Context7 MCP Tutorial: Get Instant RAG for Your AI Coders
Context7 MCP Tutorial: Get Instant RAG for Your AI Coders – In this video, we dive into the power of the new Context7 MCP and how it dramatically improves AI coding workflows by delivering precise, real-time context to tools like Cursor and Windsurf. You’ll see how the Context7 Cline MCP AI Coder avoids common issues caused by outdated model...
Which is kind of related to the very topic I'm asking about here. Although since CSS isn't a Framework exactly In the sense that React is I don't know that css is documented for migration from old code to new code to help AIs. We've come full circle
By the way, I watch Kevin's videos all the time. I think they're great. He's always getting excited about the possibilities of new CSS and when that CSS becomes Baseline. I find it interesting that this chat room doesn't think similar things aren't happening in the AI field that are happening in CSS. Css used to be very frustrating years ago, But now CSS is good and getting better all the time... AI is improving too al the time. Yall seem to be stuck on what it is today.
what I'm stuck on is mostly the reprehensible energy use and indefensible training practices, tbh
good, means you have morals
I don't see those points changing any time soon, so I will continue to avoid AI. The added benefit of still knowing how to code without is just a bonus 🤷
But I don’t. I’m goods enough a dev to RTFM and do it myself. Trying to get good, workable, succinct code from an LLM just slows me down
ah, gotcha. thanks, i think I understand your perspectives now... you DO NOT like AI tools. Fair enough.
Maybe my initial post wasn't explicit enough. It was only implied I guess that I was seeking AI tool suggestions from people who like and use AI tools. I think I'm going to repost this making that explicit.
No need to repost. The advice will be the same: LLMs can’t use new features until they’ve been trained on them. Meaning they’ll be months behind and will always favour old, outdated, ways because they have more of it in their training data
i don't know that you're a authoritative enough on AI to make that call. You don't know what MCP or RAG are.
"not knowing about MCP and RAG for AI"
is equivalent to
"not knowing how to center a div in CSS"
Doesn’t change how LLMs work
The way LLMs work is changing all the time.
memory, reasoning, diffusion, tooling, etc...
True. However, the fact that it’s giving you the problems you’re having says that my point still stands
I'm glad you weren't on the original CSS Working Group. You seem stuck on what is.
considering you opened a new thread, I'm locking this one. There's no further point in trying to convince each other.
me: "how do we try to resolve problem A?"
you: "Problem A, Problem A, Problem A!"
lock it up
👍