Context Engineering
Change a few lines and the same model behaves like a different one; what shifted is rarely the model but what it was shown this turn and in what order.
Context engineering works on four things: selection, ordering, compression and isolation. This is not prompt engineering: prompts care how you ask, context engineering cares what the model can see this time. When retrieval Q&A answers beside the point, the cause is rarely a weak model but wrong chunking or muddled ranking.
