Notes on vibe coding 3 - what works for me.

I’ve been doing a ton of vibe coding. A ton. I have a fairly large app that I’ve built with significant functionality. For my money it is as production ready as many things out there.

I've been really following two modes - make it as robust as possible vs just make the thing already.

For making it as robust as possible I have some patterns that I'm finding productive.

Lots of automated tests, in particular for API endpoints, and views. The larger app I have running now has over 380 pytest tests, all passing.
Creating design documents in markdown.
When a knotty problem comes up that clause can’t solve I get it to write up what it’s tried, and what it thinks the problem is. Then I send that to a different LLM. Then I send that report back to Claude and that usually helps.
git. Always git.
I use warp as my terminal and it has a chat mode. Warp's chat mode can write fantastic commit messages. I use that.
For my larger projects, I get Claude to write user facing api docs. Claude then uses those docs when creating new features.
Regularly pruning. I'll get Claude to review the codebase with strict instructions to look for duplication and anti patterns. This works well.
I setup a docker container for running dangerously skip permissions for more fun throw away projects. That's been a blast.

One thing I’ve been reflecting on is the enhance sense of FONO - fear of not Operating. The damn things can do so much and I have so many ideas and yet I don’t have an agent that can just head off and work on things for me. So I feel like I should be making time every day to do attend to Claude code, but I don’t have that time. That will change as modalities of agents improve but it’s an interesting moment and an interesting sense of time slipping through my fingers.

GPT-5 is also a beast.