# Jessica — Kitchen Mode Refinement **Status:** Designed, not yet built. Captured 2026-05-30 after a live dinner-party cook. **Owner context:** Maitland cooked a full grilled dinner using Jessica as a hands-free sous-chef. The menu/timeline worked; the *interaction model* needs work. ## Problem (from the live retro) Two complaints, one root cause: 1. **Mic is always on.** In a kitchen with multiple people talking, ambient conversation gets picked up and muddles her. 2. **She reads too many steps at once.** Partly recipe formatting (compound steps — one numbered item held 5 actions), partly behavior (she keeps going). **Root cause:** Jessica behaves like a *continuous, autonomous narrator* when a live, social, hands-busy kitchen needs a *turn-based, addressed-only sous-chef* — speaks only when addressed, advances only when told, never runs ahead. ## Decisions (agreed) | Dimension | Decision | |-----------|----------| | **Listening** | **Wake word "Hey Jessica"** to engage, **+ a short follow-up listen-window** after she speaks so you can say "next"/"details"/"yes" without re-waking her; she re-arms the wake word when you go quiet. Wake word starts the *conversation*, not every sentence. | | **Step detail** | **Action, then offer detail.** She reads the terse action and stops; you say "details" to hear quantities/temps/why. | | **Visual** | **Voice + screen.** Big current-step card + live timers on the propped phone; advance by voice OR tap. Screen is the source of truth when the room's loud. | | **Timers** | **Ask first.** If a step carries a timer, she offers it and waits for "yes." | | **Interim mic fix** | Ship **tap-to-talk toggle first** (trivial, reliable), layer wake word on top once tuned. Deferred to implementer's judgment. | ## The keystone: a structured "step" object Three of the four decisions collapse into one change. A recipe step becomes a small structured unit instead of a sentence: ``` action: "Put the squash and corn on the grill." detail: "~475°F, 14 min, turn halfway." timer: { label: "squash", seconds: 840 } ``` This single convention powers: - **Pacing** — read `action`, hard-stop; "details" reads `detail`. - **Ask-first timers** — if `timer` present, offer it and wait for yes. - **Visual card** — render `action` big, `detail` on tap, one-tap timer button. It also fixes the "reads a paragraph" problem and improves `/pa` retrieval. **Start here.** ### Proposed markdown encoding (human-readable, backward-compatible) Keep the numbered action line; attach optional indented sub-lines. Steps without sub-lines still parse exactly as today. ```markdown ## Steps 1. Put the squash and corn on the grill. - detail: ~475°F, ~14 min, turn halfway to char. - timer: squash, 14 min 2. Make the Creamy Oregano Dressing with an immersion blender. - detail: Reserve half for serving; use half as the chicken marinade. ``` Parser rule: a numbered line = `action`; following indented `- detail:` / `- timer:` lines attach to it. `timer` value is `