CHAPTER 18 · Reference Architecture and a Design Checklist · 5 / 15
Context & prompts
- Is context constructed per call and curated to a budget, not stuffed? (Ch 5)
- Do you reference content via tools rather than embedding it, with short stable handles? (Ch 5, 7)
- Do you distill/cap large tool and API payloads before they hit context? (Ch 5, 7)
- Is there prior-turn memory and a stale-content discipline (re-fetch mutable data)? (Ch 5)
- Does the system prompt carry only invariant behavior, with task instructions loaded on demand? (Ch 6)
- Are there emitted protocols (e.g. citations) the model produces and your code parses deterministically? (Ch 6)