[AgentMatrixAI]

Prompt quality and knowing your domain is critical. One issue I had early on was experimenting with LLMs to generate a frontend application in a brand new framework I was unfamiliar with (Svelte at the time) which lead to situations where I would cruise along and get stuck in a loop. The other issue came from the increasing context size that led to more unpredictable behaviors (i would ask it to change the color of a button and it would completely change the entire page).

almost all the tools i've used to date for designing frontend framework, none really replaces using cursor and being able to dive deep, however cline does seem to have gotten significantly better.

the day where you can come back to a fully working web app with moderate complexity after cleaning the gutter is still some way off but thats the dream

[bandrami]

How much of this is a cage we've built for ourselves? When I made my first website in 1998 even if we had had LLMs asking one how to change the color of a button would be ludicrous because I'd change an HTML attribute (did CSS exist back then? don't remember) either in a static file or in a perl script. That's it. And the modern "throw a hundred and seven frameworks at it" stack doesn't really do anything that simple CGI CRUD app couldn't do.

[DecoySalamander]

What has changed over the years is our expectations and requirements. Sure, you can skip "all the noise" and slap an attribute on your button, but then you'll have to track down another 20 buttons in your codebase that need a color update. You can be smart with your approaches, but at a certain project size and complexity, frameworks are unavoidable. The only choice you really have is to learn an established one or handcraft something new for every thing you make.

[energy123]

Same experience with context size and prompt quality.

Some other things I picked up:

- If you formulate a good prompt with small (but sufficient) context and it still makes mistakes after one attempt to feed the error message back to it, it's probably not going to be able to ever get it, no matter how many iterations you do with it. It will get stuck in a rut forever. Better not to argue with it.

- o1-2024-12-17 is genuinely a big step change.

[Mtinie]

When I catch an agent falling into a “debug loop” I ask it to summarize its understanding of the error. Next, I ask it to shift its context to a higher level in the code and look for patterns across my files which would impact {summary of error).

It isn’t foolproof but it has a much better success rate for me than letting it spin.

[ilrwbwrkhv]

So basically you use only tab completion fill in the middle?

[m3kw9]

By that time you can have your personal humanoid robot clean your gutter. You just do nothing but prompt, what a life that’s gonna be.

[ImHereToVote]

Can't prompting be automated?