It is very difficult.

I consider myself reasonably skilled in AI usage (multiple harnesses, skills, MCP, etc). I use only the latest models: Opus 5, GPT 5.6 Sol. I experiment with different plan-build-evaluate frameworks like GSD, OpenSpec, Superpowers. I have a reasonable AGENTS.md without much cruft: test coverage ~80%, ASD-STE100 English, use uv - things like that.

I can churn out single-use scripts and micro-projects like there's no tomorrow - one shot GSD in autonomous mode usually nails it.

But when it comes to any larger software, it STARTS very promising, but very quickly becomes a quagmire of a death by thousand cuts. I steer the general ideas well enough, but the amount of code and tests quickly grows overwhelming, weird shit starts to creep in, 2000-test harness starts to test literal things, features error out, fixes take longer and produce tons of defensive code. Asking the model to "refactor if it improves readability and reduces complexity" usually only increases complexity. After $3-5k in tokens spent on a 100k+ LoC monstrosity I just don't even want to touch it anymore. AI feels like a trap that lures you with easy wins but then you pay the debts.

Many words to say the same as many others in this thread: good design takes time and working WITH the code, seeing and internalizing the decisions.

I also don't know what to do. Writing things completely by hand just doesn't feel right anymore, and co-designing with AI (as in laying out the classes and function contracts etc) feels weird because you need to iterate to build understanding, but the LLM will happily follow any stupid idea you happened to have.

Quadruple all of this when you work in a team and keep getting 5-minute effort 10k LoC PRs with 20-line load-bearing honest caveat comments.

Maybe it's time to get to woodworking.