Use LLM APIs correctly (message roles, streaming, retries, structured outputs, tool calling). Reason about tokens, context windows, latency, and cost — and budget prompts accordingly. Control ...