Claude 4's Nasty Habit

As good as Claude 4 is, it has a nasty habit of trying to "fix" failing tests by skipping or disabling them in my experience.

So much is changing about software development, but this hasn't:

Effective tests, layered appropriately are the best way to tell whether your system is doing what you want it to do.

The cost of writing any one line of code is dropping to near zero - but the value of good (augmented) software engineering on systems that do something of value is going up as a result, at least for now.

Maybe it's because my son is obsessed with dinosaurs, but I keep thinking about the raptors testing the fences - stay vigilant out there 😉

P.S. I don't actually think the AI is trying to escape and go skynet, if it wanted to the existence of --dangerously-skip-permissions would mean it's already too late 😅

Join the conversation on LinkedIn