Descriptions:
At the AI Engineer conference, Isadora Martin-Dye — owner of a 225-year-old wedding venue in Virginia and builder of the AI agent that communicates with her couples — presents a detailed architecture talk on why single-layer system prompts inevitably fail and what to build instead.
Her diagnosis: a single system prompt is asked to be immutable, situational, expressive, and self-checking simultaneously — four incompatible jobs that cause it to handle the middle cases reasonably while breaking on the edges. Her solution is a four-layer “prompt stack” assembled in a fixed order every time. Layer 1 holds hard brand identity rules that nothing downstream can override. Layer 2 injects real-time signals — who the user is, what they’re experiencing — before the model forms its response. Layer 3 provides example-anchored voice and tone guidance. Layer 4 runs a cheap post-generation veto pass to catch what the first three miss. Crucially, she explains why the ordering is load-bearing: rendering soft emotional context before numeric constraints produces more natural prose, because the model sets tone first rather than mechanically slotting qualitative language into an already-committed numeric frame.
Martin-Dye draws from real multi-deployment experience — the wedding venue agent, a personal AI companion app, and a public utility for families of missing persons — and describes replacing 24 scattered system prompts with a single canonical assembly entry point. The talk is one of the more grounded prompt engineering presentations to come out of the AI Engineer conference circuit.
📺 Source: AI Engineer · Published June 26, 2026
🏷️ Format: Keynote Launch
