GuideIntermediate

Context Engineering

This guide treats a model call's context window for what it actually is: a finite resource to budget, not a canvas for writing the perfect prompt. The driving case is Reservo's Support Assistant, a question-answering assistant (no tools, no loop) that answers member questions about cancellations, pricing, room specs, and membership using documents that are already retrieved by a deterministic mock function — real retrieval belongs to a different guide. With that case, the guide covers the system prompt as a versioned, auditable component (never free-form writing), the structure of the payload (block order, delimiters), selecting content under a fixed token budget when there are more candidates than fit, compacting the conversation history (compressing while preserving what matters, in two stages), few-shot examples treated as just another budget line item, context rot (more input tokens degrade performance, citing real research) with isolation as a conceptual mitigation, and measuring the effect of every change by form — never semantically. All the engineering (counting, selecting, structuring, compressing, measuring) actually runs on Python 3.14; the model's response is honestly treated as a concept. Token counting always uses an order-of-magnitude estimate (`len(text) // 4`), labeled as such, never a real third-party tokenizer.