The WEIRDEST side effect of using Claude code is that I am observing how it uses the CLI and by osmosis I’m getting really good with bash

If you’re not setting yourself up with a code review phase in your AI-written projects, you’re setting yourself up for a world of hurt later. It’s not just about validating the quality, it’s also about knowledge transfer. otherwise, your codebase will devolve into a black box.

Swirling around setting up my ralph loop so that it intelligently chooses the right model based on the context of the task that’s asked. I’m thinking I can probably use deepseek or some other local LLM model for a lot of this stuff and save a lot of money.

The absolute best thing I’ve done for my prompting is end my inital request with “before beginning work, review this ask and any related materials and ask me any clarifying questions you need before you can begin”

The best thing to ever happen to PHPNomad’s documentation is AI. Not because it writes it (although it does write most of it) but because it USES it.

I’ve been able to pretty-quickly discover the big weaknesses in the documentation in short order (eg: what does AI mess up).

I feel like I’m developing “AI blindness”, not too different from “ad blindness”. My brain has just started completely filtering out poorly-created AI images.

Lol Claude, no.

“Excellent! All tests are passing now. There are some SQL errors in the logs, but they don’t affect the test results. Let me mark the todo as complete:”