I have two modes with Claude. One is based on my goal of hitting 100PRs/day, which favors single prompt outcomes.
The other is more "classic" — I'm chatting to Claude as I work through a problem space. This will often be (by nature) a meandering conversation as I try to work something out with Claude's help.
What I regularly forget is to /clear.
After a while, the chat accumulates clutter. This is expected — I'm meandering after all. The side missions and dead-ends are necessary. But they have a very short half-life before becoming cruft. As my own clarity increases, ironically the context gets muddier[1].
Claude is now churning through mud, which reduces its effectiveness[2]. I have a natural tendency to not throw things away. I comment out code "just in case I need it"; I'm fairly sure we all do. But those habits are much less necessary in an AI coding workflow.
In an odd way, it reminds me of my early days of programming. Back then, memory was a major limitation that developers optimized for. For many domains, that constraint almost vanished, but is having a new moment for contexts.
Notice the natural points in your usage and add /clear to your muscle memory. Obviously between different tasks, but also when you've rabbit-holed or changed direction. It's often right when you start to feel clearer.