Shape prompts and manage context

Chat models receive a list of messages: a system message (standing instructions from the application), then alternating user and assistant turns. Under the hood these become one long token sequence. Good prompts usually include:

• the task, stated plainly - and why, if it helps the model make judgment calls;
• the context needed, clearly separated from instructions (for example in XML-style tags);
• examples of good outputs (few-shot prompting) when the format or style matters;
• the output format, ideally a schema your code validates.

Asking a model to think step by step before answering often improves accuracy on reasoning tasks; many current models do this automatically as built-in “extended thinking”.

Context windows are big now - hundreds of thousands of tokens - but not free: every token costs money and time, and models use information in the middle of very long contexts less reliably. When history grows, summarize old turns, keep key facts (like that order number) pinned in the system prompt or a notes block, and retrieve documents instead of pasting everything.