I don’t think I’ve ever had so much fun in my programming career as I do now. Which is strange, because a few weeks ago I was in a very different place. I was watching - in horror - as the machine on my desk was taking over my craft. Like most people I guess, I derive quite a lot of my identity from that craft; hence the horror. (Let’s ignore for now whether that’s a good thing or not.) ...
So there’s this thing I’ve been mentioning to so many people recently: “Make every problem Claude’s problem”. What do I mean by that. Basically, if you have a problem. Make a folder, open claude and tell it (with /voice - Yes, I’m talking to my machines all day. Like a lunatic.): “Hey Claude, make a CLAUDE.md file with the following: You are now the owner of problem X. I expect you to first research how you can deal with it, at least for now and ideally permanently. I’ll run you once in a while, so you’ll get to check and/or re-fix and the opportunity to tune your approach.” ...
The problem wasn’t tokens. It was having to keep wondering. Every long session with Claude, you’re running a background process in your head. Is it still sharp? Am I burning too fast? What model did I switch to? You don’t notice this overhead until you eliminate it — and then you realize how much clearer everything gets. I built a status line that answers all five questions at a glance. ...
You’re probably not running out of tokens. You’re running out of confidence. If your plan is too small, no widget fixes that — upgrade first. But if you’re on a flat-rate subscription and still rationing prompts like a hiker rationing water on a trail that turned out to be two miles long, the problem isn’t supply. It’s visibility. You self-throttle because you can’t see the gauge, and you finish the day with budget to spare and nothing ambitious to show for it. That’s overhead, not output. ...
Something subtle happens as your Claude Code context fills up. The output gets slightly less sharp. The architecture suggestions get more generic. The creative leaps come less often. The code still works, but it’s more… obvious. This isn’t the old “AI forgot what I said” problem — Opus 4.6 stays on target like a bloodhound, even at 80% context. It’s more like the difference between a colleague who’s fresh in the morning and one who’s been in meetings all day. Still competent. Just not at their best. ...