Post

My AI enthusiasm ebbs and flows. Today I’m realizing that some of my perceived productivity increase is an illusion. I merely moved some of the effort from writing code to having to micromanage a coder.

Curious: How do you start a feature? Do you just describe the feature and tell it to go? Do you have it interview you? Do you write down a spec or plan? Something else?

We usually have a long discussion, plan mode, then execute. But the discussion can feel like micromanagement too because Infind I have to spell a lot more out to it to get a good plan than I would get from an experienced colleague.

And an inexperienced colleague is becoming an experienced colleague when they did the same with me. With the agent, I have to keep diagnosing why the lessons we wrote down aren’t helping them do better the next time. It’s like debugging markdown and that’s harder than debugging code!

Yeah this is something I struggle with a lot. Giving feedback to a junior engineer "sticks" while it doesn't with agents, even with memory and updated instructions and stuff it's always a crapshoot whether it'll just do all the same stuff you had to correct last time

I still see this. “Did you do <thing marked MANDATORY in guidance> first?” And get back “Honestly, no”. I’m beginning to suspect Opus 5 is particularly like this but not speaking from enough firsthand experience with other models.

Yeah opus 5 ignores my INVARIANT, REQUIRED, MANDATORY independent adversarial review step like 90% of the time

So I’m reviewing my recollection of what I tried and part of the problem is Opus is writing down what I say to write down in guidance poorly, turns out. But I don’t want to write this markdown myself. That’s even deeper down the “debugging markdown” rabbit hole.

Neither of us understands how to debug this markdown. Both of us can debug code because you get instant feedback on mistakes in code while the imperfect guidance sits there waiting for the next incident.

Yeah, these things are amazing in a lot of ways but do also feel like they're way too easy to hold wrong and are made entirely of rough edges in others

One thing that is working where I’ve taken the time to do it is to thin out skills and guidance with deterministic testable software doing most of the work. But doing that is time away from the original task, so it’s not always clear that I should and then markdown & pray wins and fails again.