dawn-fde-toolkit

Appearance

Prompting ​

The EAS voice agent's behavior is driven by one large system prompt, compiled by the Blueprint platform from modular variables and module bodies. This section documents that prompt: how a call flows through it, which parts are genuine language judgment vs deterministic routing, and the first-principles rework now underway.

Why this section exists ​

The shipped prompt is effective but large and prose-heavy. It encodes a great deal of fixed control flow (ordering rules, guard conditions, output scripts) in English, which makes edits ripple and lets rules quietly contradict each other. The daily call reviews (Daily Reports) surface the behavioral cost: rigid slot-filling that loses ready-to-book callers, a quote path with no capture, drop-off/waiter choices presented without their consequences.

In June 2026 we got the go-ahead to rework the prompt from first principles rather than keep patching it. The work is organized around three artifacts:

  • Maxims — a plain-language spec of what the agent should do, built point by point from observed calls. Each maxim names the current-prompt rule it displaces. This is the spec of record: prompt text is written to satisfy the maxims, not the other way around.
  • Tabletop sample — a representative set of real calls (drawn from the daily reports) walked turn by turn against the maxims, to find gaps without a live test.
  • Call flow + rigid/LLM split — the analysis on this page: a decision-tree view of a call with every node marked as deterministic or language-dependent, used to decide how much belongs in a harness vs the prompt.

The maxims and tabletop are live working documents in the prompt sandbox; the call-flow analysis is captured here as the stable architectural artifact.

Pages ​

  • Call Flow + Rigid/LLM Split — decision-tree of a call, color-classed by what needs the model vs what a state machine could own, and what that implies for the rework.