Great comment. I don’t do everything you say but still get very high quality code out of Opus 5 with Claude. Fable 5 can be even better but I haven’t proven it enough to be confident yet.
It may get off the rail if you’re a bit ambiguous about what you want , but that is only rarely a problem lately. We invested early in good AI instructions while still keeping the context small . We also have lots of skills the AI is instructed to use under different tasks (eg it must always span a subagent go review code, test creation skill, planning procedure etc). When all is done I just can’t believe any human could have done a better job.
