How I Vibe

May 26, 2025

Some quick thoughts on how I’m using LLMs. Everything with a grain of salt.

The LLM has a few superpowers. The one to keep top-of-mind vibe coding is that an LLM can inhale text in a way I can’t even really appreciate or understand. So my job is to throw text at it. Docs. Error messages. My own rambling (using f5 on my macbook set up as hotkey for transcription).

Best practices

1: Git EVERYTHING

Sounds obvious, took forever for me to actually do.

But just commit everything. Commit every time you get something working, and then, when the next thing you ask Claude for messes everything up you just revert and keep chugging.

2: small asks

Mentality should not be ‘I’m gonna ask the AI for what it’s capable of’ like you would with a coworker.

Mentality should be: ‘I’m gonna ask the AI for what I can quickly and easily verify’. Bonus points if you can picture in your mind the changes that will be required and then if it gives you something very different you have something to dig into.

Often this means I’m rewriting my prompts midway through as I realize I can break the ask up into two or more parts. It’s worth it.

3: Cursor: ‘Ask’ mode

Cursor changes so fast I can’t keep track of all the modes and features, but on whatever version I have there’s ‘Agent’ mode and ‘Ask’ mode. I use ask mode. Because I wanna at least see all the code as it gets put into my project.

4: Silly mantra: ‘PLENTY MORE WHERE THAT CAME FROM!’

It feels bad to toss code out. Like a waste. And in a world where code is written by humans throwing out 100 lines of code is, in a way, throwing out however much time it took to write that code. But in 2025? PLENTY MORE WHERE THAT CAME FROM!

On a personal project I’m coding up with claude, I’d rather have a small codebase that I understand but there’s things don’t work rather than a working codebase but it’s verbose and bizarre. Because sparse and broken I can fix with these tools. Verbose and bizarre we’re in trouble.

And so I try to remind myself: when in doubt, start cutting stuff.

Recent Interesting Things

1: Magic-School-Bus it

With the end of my cursor billing project approaching, I recently pulled down a few…just big famous repos. Typescript. Ethereum-Go. And then I would put files in the context, ask about them, ask if they were interesting files, ask the LLM to show me some interesting files, etc. It was fun! It felt, if I had to put words to it, kinda like being a precocious youth riding the magic school bus asking questions and getting answers.