Post

Let's ask the horses' mouth! :) Seriously I like seeing and hearing about coding workflows, I feel like my efficiency is 2 steps forward, 1.5 steps back, and I want to move to only 0.5 steps back

It's so much like being a high level manager. You need to let go and focus on the overall goal. I find it just as unsettling as when I tried to coach skills and got blank faces from half the juniors.

Agree, sometimes you get a vibe it’s understood and working the way you want. And some times the reply just smells funny and you have to restore to previous. A different sort of “Vibe” coding.

We are now working on steadily nudging the initial vibecode output closer and closer to our production conventions, so that less refactoring is needed when we move from proto into production.

Sounds like the few times I've turned on Tesla's FSD during trial periods. For me, it was more stressful than driving myself.

Fwiw: I don’t care if the code’s anywhere on the spectrum from an AI to Fabrice Bellard — if you don’t take the time to review and you own it, you’re gonna end up in a bad bad place later

working with proof assistants has gotten me used to this. they have had code search tools long before LLMs. but there i also don’t need to review the code because they’re automatically formally verified

Interesting. Maybe I just need to start using the LLM to make more rosyln analyzers. A good portion of what I seek can be done with static analysis.

This is my feeling to. Also good abstractions and TESTS 😅

Basically many programmers were just promoted as product managers.

It has legitimately been a great "head of engineering simulator" for me. I get to be extremely picky about process and procedure. Occasionally it'll lie to me. I get LLMs to check each other's work. I get to say "this is not production ready" after being told something is production ready.

Cf. OpenAI “Why LMs Hallucinate.” 70%+ calibration error rates on code generation tasks. Any distributional “intuition” is bound to fail. If we could hone intuition to predict errors, so could the LLMs. If the code matters, you need to look at it carefully.

Yes I've had the same observation. It remains frustrating when you get it wrong and you will get it wrong, though. It's difficult to navigate.

Sounds (uncomfortably) close to going from individual contributor to managing a team, if at another (lower) level.

The example was: I need a crm for this. I stopped myself from adding too much detail, to see what it came up with.

My feeling is comfort: there’s an evolving set of things I trust it with, a set of things I’ll definitely check (complexity), and just overall a slowly building Getting To Know Each Other…

Training you to watch the slave closely.

This is also a skill with people. It can be hard to delegate and trust the result. Incidentally, have you considered delegating the question of wether you need to look at a piece of code to an LLM?

Even worse when reading text that smells AI generated. You dont know where the dragons lie.

Finally, it’s not just intuition. If you spend some time going back-and-forth with an agent during the planning process, it usually becomes clear whether or not it has a grasp on what needs to happen next and how to accomplish it. If it can’t grok it, then you need to review more carefully.

i’m leaning alot on my big corpo product manager brain skill of “well that’s stupid but it works so i have to fix it later”

Agreed! But to Simon's original point, knowing where we can trust it and where we cant is part of where the skill is.

Indeed. And hopefully with time, LLM output will earn our trust too. Until then, and tomyour original point, knowing when we can / cant trust LLM output is a matter of skill.

Seems similar to becoming a manager. I know that uncomfortable feeling...

a lot of people seem to want an instruction manual to follow but I think it’s important for people to develop the intuition. very “teach someone to fish” in that way.

Learning when to trust AI output is the new developer skill. Blind trust fails, but micromanaging defeats the purpose.

If the loops had tight validation, verification & review steps builtin, I believe the output would get a lot better. Those steps could be LLM driven too, just different prompts and possibly even faster models.

it helps to do actual vibe coding on low value code, get a feel for when they’ll do well. and then build on that to higher value work i have a good sensitivity around what gpt-5 will do in Cursor vs codex-cli

It doesn't have to be like that, though: if you want to get well designed code out of an LLM you can do that by putting in extra work yourself, see simonwillison.net/2025/Oct/7/v... and mitchellh.com/writing/non-...

Control freakery?

It’s a Bluesky bug, you can’t use . notation.

It’s a hard comparison, BUT, I began trusting AI much more after AI caught more bugs in PR reviews than I did, and it did a better/faster job fixing them than the original dev. Maybe not without exception; but QA went up and my time went down.

It’s less about the machine understanding it and more making it still approachable quickly for the engineers used to a very codified way of working that’s treated them pretty well. I ask because this is my situation and I’ve been pondering about it a lot! Haha

Quite possibly, I don’t think I chose my words very carefully. I guess I’m talking about what might be called achieving “flow state”. For me at least that means 👇🏻

I feel much more out of the loop with agentic coding tools. With e.g copikot, not that much.

and on the other hand even the „good“ results of that software engineering practice is nothing that changed my world. it’s the nice to have apps that i created not one must have. might be only me but every time i try to build must have parts it’s just not good

If it can spike to your acceptance and do review, why can’t it implement?

I’ve built a workflow that generates about 10,000 words of plan for a feature. All that gets reviewed carefully. Then I feel a lot more confident letting the agent break the plan down and churn through it.

I don't think that's true: even a small bash script can take me 10-30 minutes to write by hand, compared with 1-2m with an LLM Recent example, this uv-test script til.simonwillison.net/python/uv-te... - I actually had ChatGPT build that one on the spur of the moment while writing that article