Dopamine, Discipline, and Anxiety
One Month With Claude Code
About one month ago, I switched from GitHub Copilot to a Claude Code Max 5x Subscription. These weeks were the most productive of my coding life. Then I realized I was breaking the core values I normally teach. And eventually I found myself checking my phone constantly instead of enjoying the summer at the lake.
Read on for the good, the bad, and the ugly of my Claude Code journey.
My Setup
I'm using a Claude Code Max 5x plan that costs (as of July 2026) 100$ per month. You're subject to weekly limits as well as 5-hour session limits. As the name suggests, those limits are 5x the regular Pro Plan ($20/month).
I’m using nWave.ai to develop, following a thorough process based on known good practices from software engineering. More on this in my previous post:
Still Using AI to Develop a Software Product
A year ago I wrote Using AI to Develop a Software Product, an experience report on how AI tooling has impacted the development of Lighthouse. Twelve months on, we’re still at it, but almost everything about the how has changed.
So let's dig into my learnings from the last four weeks, starting with the good.
The Good
The combination of the Opus model and nWave is such a pleasure to work with. It works the way I always wanted to work, and it's genuinely fun. If this is the future of Software Engineering, I'm ok with that!
When I started, I realized that many things are suddenly possible in a lot less time than when I would have done them by hand or with my previous workflow based on GitHub Copilot. I was rarely rate-limited; things were built fast, and built well.
And while the number of commits is certainly not a good metric to show quality, the chart below (try to guess at what point I started using Claude Code…) shows how much more I managed to do!
It felt like I could build everything I ever wanted (and more), in no time.
Solve that annoying bug that is hard to reproduce. Improve that part of the code! ALL THE FEATURES I CAN THINK OF!!
I can even do it in parallel. Why choose between new features, bug fixes, and refactorings, if we can spin up multiple sessions in parallel? Git worktree much?
I’ll be honest, that feeling I got after the first week was amazing.
Every time you manage to deliver something, you get a little (or big) high.
And you manage to deliver constantly. And you want that feeling back. So get right back at it. More features. More parallel sessions. Release new versions. Get great feedback. Repeat!
The Bad
I happen to be a trainer with ProKanban.org. We talk a lot about controlling your work in progress. About focusing on value. I genuinely believe in the things we teach.
And yet, I realized I was doing five things in parallel. Not because they were the most valuable things to do, but because I could. Read: I had tokens remaining.
But who has time to think about value if your session limit resets in two hours anyway?
I was throwing every good habit I not only follow, but get paid to teach, overboard. I'm supposed to be an expert in controlling work in progress. And here I was, with five Claude Code sessions open.
Maybe we really can increase Work in Progress without hurting Cycle Time, with the right tools and the right guardrails. But I had neither. What I had was context-switching costs, and a quiet hit to the quality of every individual piece of work.
I was hunting a quick dopamine high, finishing something fast, instead of asking what would actually be valuable for the product
And figuring out what's valuable means talking to users. But who has time to talk to humans when three agents are waiting on your input?
The one good thing about this bad behaviour: once you see it, you can fight it. I'd fallen into a new kind of utilization trap; I paid for the tokens, so I'd better use them. Naming it helped. But it didn't fix it. Something else was starting to worry me more.
The Ugly
We were at the lake, enjoying the first proper summer sun of the year. And the thing I was thinking about was…
Is my agent stuck? Let me check on my phone and instruct it - just in case…
I was becoming a lot more conscious about the 5-hour windows. Those are tokens I never get back if I don't use them. I must build more things. Interrupting dinner to quickly check what's going on. Thinking about kicking off some work more or less as the first thing in the morning.
The anxiety was real. I was aware that this is not good. I still played along. Needed the dopamine of finishing stuff. Turns out I wasn't the only one.
I never set my alarm clock early just to trigger something. Not yet. Wouldn't rule it out if I'd kept going like this. There is a great article from Steve Yegge called “The AI Vampire” about exactly this, worth a read once you're through this post.
Instead of me instructing Claude Code, it felt like I was now controlled by Anthropic and their session limits.
Claude is down? I'm free to do what I want, yay!
Oh, they're back up again, and reset session limits for all… change of plans! I can should build that big thing now…
I could feel it draining me, while at the same time I was addicted to building so much, so fast, at such high quality. Not a good combination. A recipe for disaster.
Now What?
So I did the only reasonable thing I could think of. I cancelled my Max 5x subscription and downsized to Pro. There'll be a follow-up on how that's going. But so far, I like that hitting limits faster forces me into actual breaks. As a matter of fact, I was writing this post while I was out of tokens for the current session 😉
I have a vacation coming up, and I want to be sure the only thing I worry about is where to find good food (which isn't really a problem in Italy 🍕).
I'm aware that nobody and nothing forced me into this kind of behaviour. There are probably levers a company like Anthropic could pull to prevent it. But ultimately, the reason this happened is also that they built a genuinely powerful ecosystem with their models and Claude Code. I'd recommend everyone try it, yes, even if you're not writing code day to day.
Will I move back to a higher subscription eventually? At the time of writing, I'd say yes. The value is insane. But first, I need to actually apply what I teach: control my work in progress, focus on value, and build some guardrails before I let myself back in.