Just One More Prompt

It's 11pm on a Tuesday. I finished what I set out to do an hour ago. I also have work tomorrow so I should really go to sleep.

But the last prompt gave me an idea, and that idea only needs one more prompt, and now I'm three prompts deep into something I hadn't planned for.

The laptop is still open and my brain is still lit up. Just one more.

If you've obsessively played Civilization, you know this feeling. "Just one more turn" at midnight becomes 4am before you notice. As a teenager I lost entire nights to that loop. Today's version is the LLM prompt cycle — except: it feels productive, which makes it even harder to stop.

The LLM Slot Machine

A colleague compared it to a slot machine. Every prompt you send comes back with results and two new ideas. The tree branches. Just like in operant conditioning, a nice thing might come back ... or not.

LLMs are nondeterministic and can feel a bit unpredictable at times — that's part of what can drive these kinds of behavioral loops. It's the variable ratio schedule that can make slot machines addictive.

It's also so easy to start things now — absurdly easy. The first 80% of a task used to take 80% of the effort. Now it takes almost nothing. So you start another thing. And another. Your work-in-progress explodes.

This is exhausting and addictive and genuinely fun, all at once.

For me, all of this comes in waves of a few weeks. When I'm in such a wave, my brain is constantly on. There's no downtime between hard decisions anymore — that "spa for the brain" I described in Agentic Coding Compresses Cognitive Effort is gone.

What's left is a stream of decisions, one after another, with a dopamine hit every time the agent comes back with something that works.

Yep, it's a video game. But instead of earning points, often enough you actually build working software.

The doomscrolling parallel is hard to ignore. You know you should stop. You keep going. But doomscrolling gives you nothing — no accomplishment, no artifact. Agentic coding gives you real output. That makes it more seductive, not less.

And it feeds on itself. Every "just one more prompt" round widens the gap between what your system does and what you actually understand about it. That's cognitive debt.

Then you step away — even just overnight — and coming back is painful. The mental model you never built isn't there to welcome you back. So what do you do? You start something new. Because starting is free now.

The slot machine causes drift, drift makes re-engagement hard, and hard re-engagement pushes you toward yet another new thing. It's a cycle.

Is This Still Fun Or Is It Compulsion?

I don't think the answer is discipline in the way we usually mean it — gritting your teeth, setting timers, forcing yourself to stop. It's more like re-learning to listen to yourself.

Noticing when the fun has tipped into compulsion. Feeling how you actually feel instead of riding the next dopamine wave. Working and experimenting with LLMs benefits from strong metacognition more than from discipline.

Some of my best work has come from phases like this — manic focus, total ownership, ideas flowing. But it doesn't last. It can't. What I'm trying to figure out is the rhythm: when to ride the wave, when to step off. How to finish my day with a win and actually feel that it's enough.

It might be a good idea to never empty the well – to stop when I still know what I'd do next so that getting back into it the next day is easy and gives me immediate momentum. At the same time, stopping also means that I have to let go of my own context for the moment, only to have to build it up again the next time I engage with this project. That also has a real cost and a probability of going wrong.

I don't know what the best way to handle this is. It feels obvious that the stopping skill matters more than the starting skill now.

I also notice two aspects with it: emotionally, we have to practice letting go for the day. And cognitively, we need to get better at saving context at the end of a session and resuming when we pick things up again.

Optimize For Learning Rate

The important part here is, in my opinion, to try out different things regularly. Keep the practices that work, ditch the others, and constantly evaluate.

I made a physical Kanban board for myself that makes my work-in-progress visible. It's a small thing but it's already improved things for me.

Yet, I closed the laptop at midnight last Tuesday and immediately thought of one more thing I could try.

This is, by the way, exactly the kind of phenomenon we're interested in at LESE Lab: what it feels like to make software and how that's changing.