Context

• A long chat gets slower, pricier and starts contradicting itself: the window is filling up.
• A big PDF "was not read properly": it did not fit, or the important part was in the middle.
• An agent loses track halfway through a task: too many tool results piled up.
• The same question gives different answers in different chats: the context differed, not the model.

1. Put the important things first and last. Models pay the most attention to the start and the end of the context.
2. Give instructions once, in the system prompt, not scattered through the chat.
3. Summarize before you continue. After a long discussion, ask for a summary, start a new chat, paste the summary.
4. Attach, do not paste, long material; let the tool chunk it. For very long material, use a knowledge base so only the relevant pages enter the context.
5. Trim tool output in agents: keep the fields you need, drop the rest before the next step.

• One topic per chat. New topic, new chat.
• Before a complex question, paste the two or three facts it depends on, even if you said them earlier.
• Ask the model what it currently knows about the task; the answer shows what is in its context.
• In automations, log the full prompt once in a while and read it; you will find stale instructions and duplicated text.
• Use the provider's prompt caching for the parts that repeat on every call; it is cheaper and faster.

• A bigger window is not free: every token in it is billed on every turn, and very full windows answer worse, not better.
• "It read the whole document" usually means it skimmed. Ask about a specific section to check.
• Instructions hidden inside attached content (a web page, an email) land in the context like anything else. Tell the model to treat attachments as data.
• Different models have different windows; a prompt that fits one may not fit another.

Context is the size of the desk; memory is what you keep in the drawer; the knowledge base is the filing cabinet.