We are also working on tools for AI for svelte...we will likely build an official MCP server that should help and direct AI not only with docs but also with fixers to signal various issues to the llm through static analysis.
We are also working on tools for AI for svelte...we will likely build an official MCP server that should help and direct AI not only with docs but also with fixers to signal various issues to the llm through static analysis.
Appreciate the MCP server (hat tip too to@khromov.se's). Though I haven't had much joy yet with cheap LLM + Aider + MCP docs. So static analysis sounds an order of magnitude better. I think a simple thing (given LLMs are pattern-matchers not reasoners) is just lots of good examples.
Yeah one of the problems with the MCP is that the AI needs to be smart enough to use it and even to know how to use it, especially if it's just looking up docs. @khromov.se recently added the docs also as resources so if you need your task needs a specific portion of the svelte docs you can also
include that by default. I can't wait to implement the auto fixers because I think we can really make good use of our expertise with svelte to guide the LLM into writing good code. The idea would be that the LLM would send the code to the MCP and we return a list of tips to make it better.
One hundred million percent. $effect is the perfect starting point for this. (also: ideally not tips, just examples -- LLMs can't reason! they're just pattern matchers!)
1/2 the more I think about this, it's a good example of programming shifting: - old way: fuck around and find out (it's obvious when your code is wrong) - LLMs: think in terms of tests, because that seemingly-correct linux kernel code (fucking around) probably breaks later (finding out...too late).
2/2 I think Rich's attitude towards LLMs contrasts with the attitude that first got me into Svelte. LLMs open up programming. The 'robot holding your hand' is a *good* thing for new coders. Maybe I'm wrong, but I think a positive direction is possible here.
It is definitely a good thing...however at least for the moment it needs to be approached with the right amount of skepticism. I use ai and I've had success with it but that's because I have experience and I can recognize a bad pattern. A robot holding your hand while you learn but writing bad code
Yeah, I agree with this. Avaiation has a similar concept of new pilots flying with automation and a glass cockpit. Automation but use it too early it can inhibit your path the mastering the skill in the first place.
is not what a newcomer should have to it's side. Writing code that might seem to work until it doesn't it's the worst possible start for a student. And writing tests is a non solution because that also requires expertise. This might be solved in 1 year but for now we are not there yet imho
Completely agree. Hence why my first post is that people*already* good at react are good at vibe coding. So again there is space for Svelte to carve the third way you describe.
In the same way the Tutorial says "Don't copy-paste", it makes sense to say "Don't LLM". But how do we avoid losing the spirit of 2019?
`*adds TODO 'raid paolo's github for examples'*`