Quick tip: Claude Code can hold a million tokens of context. I almost never let it get close. I keep my sessions under 30 percent. I will tolerate 40. Very rarely 50. After that I compact or clear and start fresh. Here is why. The bigger the context gets, the more tokens every single command burns, and the worse the output gets. At 29 percent, every command drags that whole 29 percent along with it. You pay for all of it, every time. That is slower and dumber on every step. So a bloated session costs you twice: more money and worse work. When a session starts to fill, I do not push through it. I compact, keep the important stuff, and keep moving. Lean sessions think better.