Those bits and more are some of the thoughts I wanted to get down based on a bunch of AI coding I’ve been doing.

More at:
www.bricoleur.org/2025/05/some...

4/4

And, LLMs opacity + non-deterministicness make evaluation (including human) even more important, as Lili Jiang effectively argues.

3/4

One-shots & waterfall encourage thinking of software as products not services, which is not good.

2/4

That seems wrong and typical Claude ("I found a test I don't pass, I'm going to pass by removing the test") but I don't know bandit, so I look online and find out that it is a perfectly smart thing to do and that bandit shouldn't check tests.

See bandit.readthedocs.io/en/1.7.3/con...

2/

Thank you for your leadership, courage, and example!

... What then, is the responsible course of action? For me, the answer now lies in refusal, the withdrawal of participation from systems that require dishonesty as the price of belonging."

time.com/7285045/resi...

2/

7) this type of coding is a bit like social media scrolling in terms of dopamine slot machine (someone at Coding AI www.oreilly.com/CodingwithAI... said this and I agree but forgot who said it)

4/

"yes the tests are important, you should still do the tests and not move on if some are failing.";
5) pay attention to Claude Code and intervene;
6) Claude Code will do better in areas that you know because you'll be able to tell when it is not doing good stuff and stop/redirect it;

3/

3) anything you would want to have at your disposal when coding, make sure Claude Code has and knows it has;
4) stop Claude Code often to point out obvious things -- "that is out of scope for this step", "mocking the test result doesn't mean you passed the test",

2/

I think that's directionally right, but I'm loathe to generalize. I have been doing a bunch of programming with the models and they often are quite good on judgment-like stuff and especially so when I know what I'm doing in the area.

I do think that a lot of what passes for AI "intelligence" these days is really AI producing a bunch of results and skilled humans choosing whether it is worth using / done / needs more prompting to get a better answer. Deep knowledge can be extremely helpful for the human part of that equation.

OK, learning git wasn't so bad. I used this video by Gwendolyn Faraday, which starts very basic but gets the job done:
www.youtube.com/watch?v=RGOj...
then this one on pre-commit:
www.youtube.com/watch?v=nX6w...

Ironically (?), I want to learn this stuff so that I can feel comfortable letting an AI coding helper remember it for me.