On Sunday evening at 19:21 I ran out of Claude Code tokens. The limit reset on Tuesday at 2:59 in the morning. That is just under 32 hours, with one Monday in it. It had been a long time since I went without for more than a few hours.

When I dictated the first version of this note, on that Monday, I was convinced it had started on Saturday evening and that I had lost two full days. I only found out it was wrong when I looked up the message I had sent a colleague at the moment the meter hit 100%.

Claude Code came out as a research preview on February 24, 2025, and I started using it in the first week or so. That is close to 600 days ago.

A day of nothing

I produced nothing worthwhile that Monday. Doing things by hand felt like a waste of time when the same work would take a few minutes on Tuesday morning. I kept a list of things to hand over once the tokens were back.

I did spend half an hour getting a local model to run on my MacBook. I recently got a bigger machine and bigger models fit on it. One runs now. It is small, and nowhere near a substitute for Opus 5.5. I tried because my dependency bothers me twice over. I depend on language models for nearly everything I make, and the ones I depend on belong to a US company.

The hum

What I had not expected is that my head got quiet. Only then did I notice how much of my attention normally goes to whatever is running. Every agent has a thread on a computer and a thread in my mind. How far did it get? What did it do? That goes on when I am doing something entirely different, in the evening or over a weekend. Some part of me stays with the sessions.

There have been plenty of moments in the past year and a half when nothing was running, and the hum did not stop then. So it is also about what could be running. An idle Claude Code is an opportunity cost, and apparently my brain keeps the tally all day. Without tokens it had nothing to count.

I had a clarity that day that I have not had for a while. That it took an empty token budget to get there disturbs me.

Nowhere to dictate

This note started as a text file, which I had never done before. For about a year, writing a week note has meant opening a Claude Code session and rambling into it through Wispr Flow. Unfinished thoughts go in, mixed with strong opinions, dictation mistakes and instructions like “look that up” or “fact check that”. I tell it to ask me questions and to make suggestions. Then the two of us massage the ramble into a document.

Without tokens I did not know where to go with a thought. I ended up dictating into a plain file on my machine, so I could feed it in as a prompt on Tuesday. No questions came back, and when I dictated it I had no idea what would come out. On Tuesday morning I fed it in. You are reading the result, after the usual back and forth.

Claude Code changed how I produce anything, and code was only the first thing. When it is gone I am not inclined to produce at all, because working without it no longer fits how I work.

Eight hours

This touches a question that is coming up at work: what do we expect from engineers now that they have this machine? It raises the bar, obviously. But an engineer’s day used to be a mixture of intellectual work and boring work, and the boring work is what the machines take. How much purely intellectual work fits in a day? Can someone do eight or nine hours of it? I don’t know, and I have not heard anyone at work answer it either.

The hum comes on top of that. If every running agent keeps a thread occupied in an engineer’s head, that thread is also occupied in the evening and at three in the morning, outside any hours anyone counts. I wonder what we are doing to ourselves. I don’t think it is sustainable.

2:59

I love working with Claude Code, and I have never been this productive. I make more useful things than I ever did, often while doing something else, which was certainly not true before. But it affects my thinking and how clear my head is, and I have not found a way to deal with that.

On Tuesday I woke up just after three. The tokens had come back at 2:59. I considered kicking off the tasks that were waiting and came close, but I went back to sleep. At 6:30 I started them, before my first coffee.