Seeing the final result of a feature doesn't really give me any dopamine. Maybe because I'm mostly working on projects I know how to do. When I give it a prompt I already know what the result should look like, so I'm not really surprised by anything it produces.

I found Claude extremely addicting at first (the dopamine hits were real for me!) but over time I guess I've gotten desensitized.

For me it's always magic to see it work, even if it's a tiny change I'd need to do. To be able to ask "add an opt-in flag for this part of the script" and see it work, updating documentation and asking follow-up questions when instructions are vague it's impressive...

The "uncertain reward" nature of LLM usage makes it a skinner box, yes.

Everything can be optimized, performance can be improved, you can always think of more edge cases and user stories to cover everything, but after a point that just becomes procrastination in the form of chasing perfection. It's also hell if you've got even the slightest bit of ADHD, rapidly leading to task paralysis with the sheer scale of the plan.

Now I sit with a notebook sketch out everything I am thinking about and then condense it to a planning prompt and then once the plan aligns with my representation of the task, I start implementing.

Yikes. I feel seen.

its very personal if its good or bad i suppose. (not a psychologist so honestly dont know if its really similar. just expressing my personal feeling about it)

I think there is kind of a meme going around about multitaskers doing very well with vibe coding, and I can see it. Although, as someone who has the opposite problem, it can be tiring if I try to do more than two things at once.

This is an example of a session shared by Simon Willison in one of his recent blog posts if that helps?

What I think I'll do is to formalize my workflow in the form of a meta project and do write-up about it. I think it's going to be more useful to a wider audience than share sessions tied to a specific niche project.

However, in a nutshell, I do a discovery session in which I explore the problem. The output of this discovery session is a work-item document with lots of tasks in it but roughly defined. I then open another refinement session, so I intentionally lose the discovery context. This session improves the work-item doc. Then I open another execution session in which I iterate over each task. Then I do a post session in which the agent is instructed to verify if the work-item is indeed concluded or if we left work out by mistake. The output of this last session is either another follow-up work item or a decision change documentation, either in the work item I'm closing or the next work item if it's already there, so I document decision changes that I needed to make during the execution phase.

They always end up praising me for the high quality code and howdthey found exactly ZERO manifest bugs in the code, and this must be the work of a skilled senior developer owing to the code's polish.

Then I point out the bug I had just discovered.... "you're exactly right!"

One of his points was it's not really the author's fault he's been sucked into believing he made something revolutionary, AIs suck up to you and constantly tell you your work is brilliant and of course it's revolutionary.

https://youtu.be/Q6nem-F8AG8?si=pQm76DpJOhVFJuW7

It will look nothing like those things, but it will be obvious in retrospect.

For better and worse.

If I had to describe the way I feel about these tools, well, I don't. They're tools. They exist, briefly, to perform a function and then ascend to whatever plane lost and forgotten programs go to when they leave your hard drive.

I feel exactly the same way about AI agents. It's a tool. I put text in, computering happens, and code comes out. The amount of effort and complexity involved is different, but fundamentally I view it as the same boring codegen we've been doing for decades.

I highly suspect that because I treat AI agents this way, I get actually good, well-engineered results. I've seen extremely low rates of bad code, inappropriate solutions, or the general slop we see in vibecoded projects. I put in a great deal of engineering effort up front, and I get engineering out.

That some people find this loop addictive is... incredibly disturbing. I only foresee very bad things coming from this.

My coding refinement sessions with codex has been just as much interesting as the implementation sessions. I get to go over the code with the agent, learn a lot of interesting techniques along the way, and we end up with nice tame code.

- Working later because of addictive loops.

- Wasting time and money building stuff nobody will ever use.

- Abusive screen time.

- Loss of deep understanding of a code base and over-reliance on the AI, with reduced incident or bug response time.

- Producing a lot of sloppy code, too much to be owned or reviewed.

- Mental fatigue.

It's MMO all over again. And I'm not complaining, but I want to raise awareness for me and other that maybe this isn't such an innocent activity as we think. Coding with agents might just be too addictive for some.

  This is how others feel as well and how software engineering will feel for new generations

How I work today is still very similar to how I worked in 2023, but now I'm typing a TON of English and very little Ruby. But the overall vibe is nicer and starting on difficult stuff is significantly easier.

Coding to me has always been the work of a craftsman. I spent years learning syntax, studying APIs, and generally becoming an expert in a language I could then use to exactly express myself. My code became my way of of telling a story. I picked that language not only because it was the right tool for the job, but because I liked the community, I liked the direction of the tool, and I liked writing it.

Agents obliterate all of that. Before, source code was like exploring a canyon. Each function, loop, and nested brackets was someone's choice. There was an implicit story to follow. Agentic code is that valley after the glaciers have retreated. A good illustration of what I'm afraid of losing is in those videos cataloging the [the rapidly dwindling sanity of valve programmers as expressed through code comments](https://www.youtube.com/watch?v=k238XpMMn38). When AI generates your code, you lose that context. You lose that human connection to code that I care so much about.

Unfortunately, I don't think the human touch was as high-of-value to other developers as I previously thought. I've realized that for most developers, it was never about the journey, it was about getting to the destination as quickly as possible. We were infected with a plague of idea-guys and didn't even know it.

That's okay.

There are a billion things in the world where what was once the sole dominion of craftsmen is now shared with machines. Tables and chairs used to be all handmade, now I can order one online and it comes in a self-assembly kit. For 99% of people, code is not the objective. The tool they can use is. They don't understand how it works, or what choices you made. It's story would be lost on them anyway, and that's okay. They just need the self-assembly kit.

I use AI when it makes sense. Writing tests, refactoring big changes, reviewing pull requests, and generating the CUD in CRUD after I define the data and how to get it. Stuff I'd send to my junior engineering assistant. I now get to spend more time working on what I want to work on. I just can't forget that learning is effort, and if I skip all of that, I'm no better than anyone else with access to Claude.

I hear a lot of developers coping with AI by saying that what makes them unique is their ideas or understanding of systems. Does anyone really believe that AI will not get cheaper, smaller, smarter, and easier to run locally? It will come for you, and understand your systems better than you.

It's all a game of balance and incentives. If you want to understand code, you need to write it. If you don't care about code, and just want what it can give you, generate it and realize that you're atrophying a unique skill. Feeling hopeless about all of this? Outside of work, nobody is making you be productive. You can just write code to enjoy it. Anyone who tells you differently is getting off at a different station.