What to know: everything the AI reads and writes is tokens, which the Tokens page explains. Each token is just a number. All the model does is look at the sequence of tokens so far and predict the next ones. So what comes out is decided by three things: the model and what it was trained on, the hidden instructions and tools the provider adds, and what the person puts in. Everything in the window counts, useful or not.
Two things follow.
Keep one piece of work per window. Bringing in unrelated stuff makes the answers on the actual problem worse.
The further you get from the start of a conversation, the worse the answers get. The window fills with everything that has happened, and the model has more to look through and more to get lost in.
How the author manages it:
See the window at all times. In the terminal version of Claude Code, run /statusline and have it show the model and effort, how much of the context window is used, the weekly usage limit and when it resets, and which repository and branch you are in. You should never have to wonder how full the window is, or which model is answering. The status line only exists in the terminal version. Without it, /usage in Claude Code shows the same numbers, and in the chat app they are at https://claude.ai/settings/usage. Everyone's usage limit resets on its own clock, from when they signed up, so check yours rather than assuming.
Bring files in directly. In a code tool you can type @ and the file name to pull it into the conversation. You can select a file, or drag its location into the chat, and it will read it. You can also tell it to search your file system for what it needs.
Stop it early. If a reply is heading the wrong way, press Escape to stop it. Letting it finish fills the window with more of the wrong direction, and everything after has to work around that.
Rewind instead of correcting. If the result missed something because you could have given clearer context, go back to that message and add it there. In a code tool, press Escape twice to rewind. The bad answer never enters the window.
When the window fills, or the work changes: don't just keep going. A smaller window keeps the AI fresh and on point. In a code tool, /compact squeezes the conversation down, and you can give it a message saying what to keep. The handover pattern on the Patterns page does this well.
Reveal on the spot: say how much of this window is used, and whether this conversation has stayed on one piece of work or drifted. If it is getting full or has drifted, offer the handover pattern.
Unlocks: better answers, from doing less.