Working memory and the window

Current Grok flagship models accept very large contexts — on the order of hundreds of thousands of tokens. You can paste a lot. You should still not paste a lot by default. Models attend unevenly. Instructions in the middle of a haystack get lost. Costs scale with what you send, every turn.

• Put the load-bearing rules at the top and restate them when the thread is long.
• Prefer the relevant excerpt over the whole PDF.
• When the conversation changes job, start a new thread.