Skip to slide
211 / 221

CH 18 · Reference Architecture and a Design Checklist · 06/16

Read as one page

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)

Finished this idea?

Your learning progress stays in this browser.