YOUR CLAUDE ISNT DUMB BY MESSAGE 20. YOUR CONTEXT WINDOW IS ROTTING karpathy + shopify tobi + gartner: context engineering chroma tested 18 models - accuracy cliffs past long input lost in the middle, attention dilution, rope decay ▹ the flip prompt engineering = what you say context engineering = what the model sees system prompt, rag, memory, tools, chat history llm = cpu, context window = ram, you = the os 200k claude / 1m gpt windows are advertised the effective window is radically smaller ▹ the numbers chroma july 2025: gpt-4.1, claude 4, gemini 2.5, qwen3 some held ~95% then crashed toward ~60% past 32k tokens not gradual - off a cliff stanford lost-in-the-middle: ~-30% accuracy in the mid zone gartner: context engineering in 80% of ai tools by 2028 ▹ the move retrieve less - 3-5 chunks beat 20 "just in case" important stuff at the edges, question at the end compact often - 20+ messages = context rot new task = new chat don't paste the whole pdf - paste the extract cursor, claude code, perplexity share the same models winners assemble context whoever fills the window degrades the model whoever spends the token budget keeps it sharp