No worries. Yeah, I have been a reluctant adopter of the technology. I honestly didn’t really buy much of the hype until Claude code and some of the better models came along.

Even frontier models screw up on their first, second, and even third tries at times. The insight that Anthropic really had was realizing that they needed to be allowed to chat with themselves in a loop, fail in a loop, and that they’re much better if you apply objective standards to the projects (things like unit tests of course, but also integration tests, static code analysis, etc).

The harness is really the thing. Even OpenCode in my experience can produce things that are fairly decent.

That’s where I think both initial implementations of code helpers (i.e., a vanilla “chat box” with “tools”, and “agentic workflows” that try to one-shot everything) that were made in the initial round of “doing AI” are effectively useless garbage that should be buried like radioactive waste.

If you don’t prod it, Claude will also spit out worthless crap that doesn’t work at all too. The key is to add all of the plumbing that reasonable projects have (code formatters, linters, etc. etc. forever) and to add gating so that Claude can cycle away getting it wrong and then correcting itself until it gets it right.

EDIT: More relatedly to this thread, I don’t think I would have any problem with some type of high school elective course where kids could use claude code or something like it to try to write some software. It’s pretty likely that there are some kids already doing so in their free time. I came of age in a similar era for technology in general, and I took three or four electives where I was doing things with computers and it was helpful. About the only argument against it I can think of is a practical one…I don’t think school IT departments are ready for the havoc the kids could potentially unleash with this stuff, and it could be that they do not have the staff necessary to support its usage.