you need to develop an eye for when your context window is actually full... and i'm not talking about the percentage the model shows you models now advertise things like 1M token windows but start degrading way earlier Claude Code is the perfect example - you're at 20-30% capacity - outputs suddenly get sloppy - the model starts forgetting key details - quality drops but the meter says you're fine this isn't something you can learn from docs or theory you have to run enough prompts to feel…