Context Window Hygiene

A bigger context window is not a memory upgrade. It is a more expensive way to be wrong.

The model attends to the prompt it actually sees. Long contexts hide signal under noise — middle-of-context recall degrades, key facts get out-ranked by chatter, and cost scales linearly with every replayed turn. Hygiene is about removing what doesn't earn its place, not adding more.

When to use this skill

• The user is running a long agentic loop (tool-use loop, multi-turn task).
• The user reports the agent "forgetting" mid-run, or contradicting earlier turns.
• The user is paying for context tokens that grow linearly with conversation length.
• The user is considering "just use a 1M-token window" as a solution to anything.

What enters context — name the categories

Every token in the prompt is one of: