Have you considered not doing that? It's not obligatory to have an LLM shit out unreviewed code for you, you're making a choice to do that, and you can make a choice not to.

Review the code. Hell, maybe even write some code yourself.

What you're describing is how I feel whenever I use an LLM for anything more than the most basic of tasks. I've gone from being a senior level software developer to managing a barely competent junior developer who's only redeeming skill is the ability to type really, really quickly. I quit the management track some time ago because I hated doing all my software development via the medium of design documents which would then be badly implemented by people who didn't care, there's no way you're going to get me to volunteer for that.

"who's only redeeming skill is the ability to type really, really quickly" This really resonates. Here’s the hard truth: AI coding assistants are like giving devs a faster keyboard. Sure, they type faster—but typing was never the bottleneck.

And while AI often produces high-quality code on the surface, when things go wrong, they take longer to fix. MTTR for AI-generated code is much higher, especially when devs don’t fully grok the code they accepted. That undercuts a lot of the perceived velocity gains, especially in complex, evolving systems.

Not OP, but you nailed my feelings perfectly. I did not like managing for precisely this reason, and it never got better. The trenches are for me.

Re LLMs I love collaborative coding because I can sometimes pick up or teach new tricks. If I'm too tired to type the boilerplate I sometimes use an LLM. These are the only two redeeming values of LLM agents: they produce code or designs I can start from when I ask them too. I rarely do.

I hope OP can find a balance that works. It's sad to see the (claimed) state of the art be a soulless crank we have to turn.

I take a hybrid approach. I will describe a simplified problem to the LLM, have it generate a well commented and reasonable approach for the problem. I then use that as a cheat sheet for implementing my actual code. This still gives me hands on codi and more control, without needing to agonize over the details of each coding technique.

I think you're on to something: It sounds like the developer needs to be more hands-on in making changes; as opposed to treating the AI like a subordinate with autonomy.

---

For example: About two years ago I worked with a contractor who was a lot more junior than our team needed. I'd give him instructions, and then the next day spend about 2-3 hours fixing his code. In total, I was spending half of my time handholding the contractor through their project.

The problem was that I had my own assignments; and the contractor was supposed to be able to do their job with minimal oversight. (IE, roughly 0.5-1.5 hours of my day.)

If the contractor was helping me with my assignment; IE, if the contractor was my assistant, I'd have loved the arrangement.

(In case you're wondering how it ended up for the contractor, we let him go and hired someone who worked out great.)

---

I suspect if the OP can figure out how to make the AI an assistant, instead of an employee with autonomy, then it will be a better arrangement. I personally haven't found an AI that does that for me, but I suspect I'm either using it incorrectly or using the wrong tools.

> managing a barely competent junior developer who's only redeeming skill is the ability to type really, really quickly

Hits the nail on the head. For an actual junior developer, they'd at least learn over time. With LLM, open up a new chat and you start with a new hire.

How large of an idea can you implement by yourself as opposed to leading a team?

I can and have lead discovery sessions either virtually or hop on plane and do it in person, design the architecture, be a halfway decent project manager, tech lead, hands on development, take an empty AWS account and set up best practices from a security, networking, deployment pipelines, etc.

But I can’t do it all at once and there is still only so much I can do at once and no business would be happy with the length of time it would take me or any one person to do a complex implementation at scale. At some point, if you have big ideas, you have to lead projects.

However LLMs make a damn good and fast junior developer that can do it all especially on a green field project with clear requirements and the amount of work I can get done now by myself on each of those levels - post requirement gathering is at least 2x - 3x.

And I am not bragging - I’m old. I should be able to do all of things I mentioned.

Code is an implementation detail. My job has always been to get computers to do stuff to either make the computer money or to save the computer money either via my own hands or via leading a team.

Maybe everyone will now expect you to complete tasks as early as possible. They might also be evaluating the time needed to finish tasks, considering the help from AI.