Context Hygiene

I was working with a builder recently who wouldn't clear their session. It had been running for days, the context was enormous, and every new message came back a little muddier than the last: old instructions bleeding into new ones, details invented that were never in the conversation. I told them to clear the session and start fresh. They pushed back. They were worried that clearing would mean losing everything, that all that accumulated context was valuable and they couldn't afford to let it go.

That worry is the tell. If you're afraid to clear a session because you might lose something important, you've already lost control of what's in it. And the harness will eventually clear it for you anyway, through some auto-compact step you don't control, on its own terms rather than yours.

Language models are entirely at the mercy of what's currently in their context. Too much of it, and the model gets distracted, treating stale or irrelevant details as if they still mattered. Too little, and it starts filling gaps with guesses, which is one place hallucination comes from. There's no way to opt out of this tradeoff. The only choice is whether you manage it or let it happen to you.

Managing it looks like a few concrete habits. Use /clear between unrelated bodies of work, so one task's context doesn't leak into the next. Before you clear, write down what was decided and what's left to do, so the next session can pick up cleanly instead of starting blind. The same habit works mid-session too: if a piece of work doesn't need to live in your main thread, like reviewing a batch of documents or running background research, hand it to a subagent and let only the summary come back.

The instinct to hold onto everything feels safe. It isn't. It's just deferring the cleanup to a moment you don't get to choose.