Context window

What fills the window

• System prompt, tool definitions, skills, attached documents, and the full conversation history.
• Every call re-sends the whole thing — that's why cost grows turn over turn.

Bigger window ≠ better

• ChatGPT launched in 2023 with a 4k window. Today's models start at ~100k and the frontier is pushing into the millions.
• 100k tokens is roughly 5 hours of human reading. The model processes that in seconds — but processing isn't understanding.
• More material competes for attention. Signal-to-noise degrades as the window fills.

Managing it

• Attach exactly what's needed, not the full document. Targeted context beats dumping everything.
• Break complex problems into checkpoints — summarize each piece in a fresh conversation, then feed those summaries forward.
• Compact or hand off when approaching the limit.