Post

Was looking into SvelteKit and got pretty far, pretty quickly in building a prototype of an app. with the help of AI. However, I also got stuck, not understanding what I had actually build. Went back to rebuilding from scratch, making sure I understand each piece of code. There is a lesson here.

1 repost9 likes

Pretty good for an LLM to get anywhere with Svelte without constant supervision Was it using Svelte 4 syntax or did it start with runes? My new favourite is correcting LLMs not to use a .server.ts if there’s a .remote.ts

No, it used Svelte 5 syntax. Also added the MCP integration for Svelte documentation. Still, being new to the space, it build it in a way that I think a seasoned Svelte dev. would never do.

It's like letting the plane fly itself until you need to take over when it's a million miles from your destination, upside down with no fuel to spare.

And I can't help but wonder; will these "cracked" teams that just started forming, cranking out way more code with AI, grind to a halt in a few months when nobody understands what they built?

I recently got a new client gig which was “we need help untangling this 80.000 lines of code. It used to work at the beginning but now it has bugs and neither we nor Claude can fix” 🙈 I did use a coding agent in the process but the lesson for me is - the human stays in the driver seat.

You’re bring the vision, the architecture, the constraints, the coding style, etc. You also have to review the agent’s output every “phase” of the way. The LLM will be happy to fill in gaps with a “robust and comprehensive” something but that’s just “what best rhymes with my training data”