Good Vibes: A Claude-Code Case-Study
8 points by surprisetalk
8 points by surprisetalk
Overall I feel this matches my experience with Claude Code, in that the better plan you have, the better output you get.
Also:
LLMs are like golfing savants – they excel at (1) driving the ball long distances and (2) putting the ball into the hole. Outside of those situations, they become lost and distraught. If your first shot doesn’t land you on “the green” (within putting distance of the hole), you should throw everything away and start over.
Fantastic analogy.
You should try listening to this with ElevenReader 😅
Be wary of many modules. It’s okay to have big files.
This is directly counter to my experience. At least roo code wants to slurp whole files and blow out its context window. But then I haven’t used Claude code in a few weeks, instead I’ve been using Cursor. And thinking about it, I’m not sure that’s been a problem with Cursor.
Yeah, my surprisingly pleasant experience with large files was one of the main reasons I wanted to share this post.
At least in Claude Code, I feel like that it spends extra context on trying to find/organize files. I think it works better when it reads entire files (~2,000 LOC) and uses its text search tool.
Any ideas on how I could objectively test this?
Make your files much bigger and much smaller? My big files are over 10kloc.
A 8k loc file at work often exceeds its limits, and it spends a bunch of tokens grepping around and reading chunks of the file instead
Yeah, 2k is about the sweetspot. Large by uncle bob standards, but mostly fine, and better than trying to memorize some arbitrary division of all of this intertwined code, and needing to refactor across files instead of within one.
But go bigger than that and Cursor starts to flail uselessly somewhat often, though generally due to the sheer length of blocks within functions. You can stretch it a bit if you split those large functions into smaller helpers. But too long a file and it pretty often just falls over.