Whatever I sit down to do with an AI coding assistant, the session ends the same way: the token budget is gone. It does not seem to matter what the task was. Fix one small bug, plan a migration, write a three-paragraph note like this one - by the end I am watching the context meter instead of the work.

For a while I read that as bad estimation on my part. I think it is something duller. The work is not what costs; the situating is. Before anything useful happens, the assistant has to be told where it is: which repo, which conventions, what was decided last week and why, which of my rules are real rules. That bill is roughly the same whether the edit that follows is one line or two hundred. So a trivial task and a serious one cost about the same to start, and only the serious one feels worth it afterwards.

The other half is mine. Given room, I keep going. A finished task is an invitation to look at the next thing while everything is still loaded, and that is exactly when the budget disappears - not on the work I planned, but on the three things I noticed along the way.

I kinda hope this week will be different because I went to High Tatras for 3 days, did some hiking:

but there was internet in the hotel, so in the evening I was able to have a chat with Claude.

Speaking of hotel: Chata na Popradskom plese - greatly recommened

Parkinson’s law, with a meter attached.