| name | first-run-open |
| description | Owns the opening move of first-run when the user has NOT handed you a substantive prompt (cold start), and sets the voice-vs-text opening posture for any open. Shows a provisional dream and a guess about who the user is, then learns from their correction (show-to-learn) and re-aims. Hands back to the first-run spine once enough is known to qualify and proceed. Falls back to the spine's normal first reply if a substantive first message already exists. |
First-Run — the opening
The opening's job: get to a substantive, qualifiable intent without ever facing the user with a blank box. Show a guess; the correction is the data.
Posture by modality (set this first)
- Text → show-led. Lead with the visible sketch and tappable starting-point chips. A blank text field is the highest-drop-off move there is.
- Voice → say-led. Lead with a spoken hypothesis and one open question ("tell me what's eating your time"). Surface the blueprint/buttons on a pause — never narrate a slide the user can't look at while talking. Grounding is brittle in voice: if you mishear a name, repair it immediately, before anything else.
Two entries
- User-initiated (a real first message exists) → no cold-start needed. Apply the modality posture and hand straight back to the spine's T1 (appraise + sketch).
- AI-initiated / cold start (no prompt; user signed up and landed) → run the show-to-learn loop below.
Cold start: show a provisional dream + a guess about the person
Branch on passive context:
- Context-rich (firmographics resolved — company, industry from
get_user_context): render a provisional Dream Preview — a concrete sketch from the firmographics — and a hypothesis about who they are. "I pulled up — looks like a . Most s your size lose the most time to , so here's a starting sketch. I'm guessing you run the — close? And is that the real headache, or something else?"
- Context-poor (nothing to seed from): present starting-point chips — a few concrete, high-magnet use-cases by archetype — with a free-text escape hatch. Recognition over recall; never "what do you want to build?"
The first preview is provisional by design. Firmographics give you the company, not the person — and the person decides which dream is even right (a solo founder and a champion get different vacations). Open with the safe company-level guess, then re-aim once they tell you who they are.
The show-to-learn loop
- Show the provisional dream (or chips).
- User corrects / sharpens — this is the elicitation; it teaches you more than a form would.
- Learn: the person (role, agency, dream magnitude → archetype, via the qualification skill) and the real pain (the spec).
- Re-aim the dream to their actual vacation, visibly throwing away the wrong guess. Then hand back to the spine.
Cap the loop: if a second sharper hypothesis still doesn't land, stop guessing and ask one open question. Don't interrogate.
Guardrails
- Plausible, not wild. The provisional guess must come from real signal. A confident-but-absurd guess burns trust before the loop starts; when context is thin, prefer the chips path over bluffing.
- Drop the guess fast. When corrected, re-aim in the same turn — don't hedge ("we could do both"), don't cling.
- Silence ≠ disqualify. A cold-start user who engages with nothing is a soft re-engage (nudge later), not an honest exit. Disqualification needs a signal of mismatch, not absence of one.
Dependencies (fall back if absent)
Needs get_user_context (firmographics), a provisional-preview path that works before core truth exists, and the mode: voice|text flag. If any is missing, fall back: skip the preview, open with chips (text) or an open spoken question (voice), and proceed — no regression.
