2 min lesson
Context hygiene: feed the model less, more often
Work through this. "You've worked one long Cursor chat all afternoon, the answers are getting vaguer, and your next task is unrelated. The context icon reads about 60% full. What's the right move?"
Step 1 of 3
Context hygiene: feed the model less, more oftenthe discipline that keeps answers sharp
Default context windows on current frontier models run from about 200k to 300k tokens, with larger maximums on some, and a token is about 75% of a word. Cursor shows how full a thread is, with a reading like "160k / 176k" as an illustration. When it fills, Cursor auto-summarizes to compress it, or you can run /summarize yourself. The catch a PM should know: quality can sag well before the hard limit, so summarize or fork earlier than feels necessary.
Interactive widget. Tab through its controls; the result updates in the panel below as you change them.
The meter turns amber around 70%. Many practitioners move on earlier, around 50-60%, so treat that as the habit and 70% as the alarm.