Lumi.
Lumi · private beta · EU · LatAm · MENA

Your second brain
for closing deals.

Speak after a showing. Forward an email. Pull up a client. Lumi captures the soft signals, fills the brief, and feeds Claude — automatically.

  • Voice → CRM, auto. No forms.
  • Works offline. Syncs when you're back.
  • Free for agents in EU · LatAm · MENA.

9-min read · Updated April 2026

9:41

Lumi · Wednesday

Good morning, Niki.

Two showings · three leads need a nudge.

Clara Ruiz
Tomorrow 11am showing at Passeig de Gràcia 84 with Clara Ruiz. She wants to bring her partner.
Got it — creating the showing.
Suggested event · 92%

Showing · Passeig de Gràcia 84

Thu · 11:00–11:45Gràcia
What’s the HOA for Apt 4?
€210 per month, covers elevator, concierge, and rooftop.DOC 12
Ask Lumi or speak…
Calendar
Todos
Lumi
Clients
Settings
agent toolkit · field guide

My phone knows when
I'm at a showing.

Most agents do 60-120 showings a quarter. Most agents log maybe 70% of them — the rest fall through the cracks of busy days, last minute reschedules, walk-by drives that turn into impromptu tours. The data lost is exactly the data that's hardest to reconstruct later. Geofence-driven capture closes the gap without the agent doing anything.

9-min readUpdated April 2026Pack 17 of 30 · @lumi.estate
geofence · murtais 24 · 14m 22s
enter:    11:02 · phone within 100m
exit:     11:16 · phone exits radius
duration: 14m 22s · engaged tier

[60s after exit, prompt fires]
  "Voice memo on Marina + Hugo at
   Murtais 24?"

[agent records 78s while walking to car]

[2m later, structured capture in CRM:]
  · soft signals: 4 captured
  · next_step: send fresh garden photo
  · decision_group: husband co-decider
  · review pending agent approval

Geofence + voice memo + Whisper + Claude = a CRM update the agent didn't have to write. The agent reviews 5 fields and approves; the briefing is captured before memory decays.

the protocol

Five steps. Setup once.

The geofence runs invisibly after permission is granted. The voice memo prompt is the only point where the agent touches the system — and they touch it for 60-90 seconds, not 6 minutes of CRM data entry.

  1. 01

    iOS geofence: 'Always' permission, 100m radius.

    iOS geofencing requires the 'Always' location permission — the agent grants this once at setup. The app registers a 100m radius around each showing's address (taken from the calendar event). When the agent's phone enters the radius during the calendar event window, the app fires a 'showing started' event. When it leaves, 'showing ended' fires with the elapsed duration. Permission is the friction; once granted, the system runs invisibly.

  2. 02

    On exit, the app prompts a voice memo.

    60 seconds after the geofence exit, the phone vibrates with a soft prompt: 'Voice memo on Marina + Hugo at Murtais 24?'. The agent taps record on the lock screen, talks for 60-90 seconds while walking to the car, taps stop. No app to open, no form to fill — the prompt arrives at the right moment and the recording happens before memory has decayed.

  3. 03

    Whisper transcribes within 30s. Claude extracts the 5 fields.

    Same pipeline as the call-log protocol — Whisper transcribes the voice memo, Claude extracts the structured capture. Use Sonnet for the soft-signal nuance; Haiku misses the unprompted observations that matter most. By the time the agent is back in the car, the structured capture is in their CRM with a 'review and approve' nudge.

  4. 04

    Agent reviews + approves. Never auto-commit.

    The structured capture surfaces in a review queue, not directly in the CRM. The agent taps through 5 fields (12 seconds), edits or rejects any that misfire, taps approve. Auto-commit feels like a CRM that's making decisions on the agent's behalf — and any miscapture (which happens ~10% of the time on noisy memos) corrupts the brief permanently. Review-then-approve preserves trust.

  5. 05

    Month-one analytics: showings logged vs. showings remembered.

    The first month of running the protocol produces a surprise: the gap between showings the agent thought they did and showings the geofence actually captured is usually 15-25%. Forgotten showings (especially second-look visits and walk-by drives that turned into impromptu showings) are now in the CRM with full context. The CRM is more accurate than memory — measurably so.

anti-patterns

Three failure modes that break trust.

Each one has been a real failure in early adopter agents. The protocol's discipline is mostly in what NOT to automate.

the auto-classifier

showing_summary: 'Brief showing of 6 minutes — client clearly not interested.'

Inferring interest from duration alone is wrong constantly. A 6-minute showing might mean the client knew within 30 seconds that the layout wasn't right (cold) or that the layout was perfect and they want to see it again with their spouse (hot). The voice memo tells you which; the duration alone doesn't.

the silent commit

[CRM update fired automatically without agent review]: 'Marina lukewarm on Murtais — auto-scheduled follow-up for Friday with similar listings.'

Auto-committed without review. If the voice memo was misheard, the brief is now wrong and the auto-scheduled follow-up will land off-target. Worse: the agent didn't approve the next step, so when Friday's email goes out and embarrasses them, they distrust the whole system.

the privacy leak

[Default settings export geofence trail to Google Maps timeline visible to family-share members]

Real failure mode: an agent's location history exposed to spouse's Google Family Sharing. Not the AI's fault, but the protocol's setup must explicitly disable family-share location export and explain the privacy boundary. Default-permissive settings turn a useful protocol into a domestic-trust accident.

the prompt that captures it

What to feed Claude.

The prompt that turns a 60-90s rambling voice memo into the 5-field structured capture. Plugs into the call-log schema for unified CRM updates across calls and showings.

arrival_system_prompt.md
You are a senior real-estate agent's
post-showing capture assistant.

INPUT
You receive: the property address, the
client present at the showing (from CRM
calendar event), the showing duration
(from geofence enter/exit timestamps),
and the agent's voice memo recorded
within 60 seconds of leaving the property.

OUTPUT
A structured capture object:

  showing_summary: <one sentence — what
                    happened, factual.>
  client_reaction: <one sentence — the
                    overall posture
                    (interested / lukewarm
                    / not for them).>
  soft_signals:    <array of 0-5 — pauses,
                    questions asked,
                    unprompted observations.>
  next_step:       <one sentence — concrete
                    follow-up with date or
                    trigger.>
  decision_group_update:
                  <one sentence — anything
                   new about who else is
                   involved.>

RULES (non-negotiable)
1. Voice memo is messy by design — the
   agent is walking back to the car. The
   prompt extracts structure; it doesn't
   require structured input.
2. Soft signals are gold — capture every
   one mentioned in the memo, plus any
   the agent referenced obliquely.
3. Showing duration is a signal:
   <8 min — almost certainly not interested
   8-25 min — engaged but not committed
   25+ min — strong interest, decision
             group probably forming
4. If the memo mentions the client by
   name, decision_group_update should
   reflect any other names mentioned.
5. Match the call-log schema (calls
   protocol) — the post-showing capture
   plugs into the same CRM brief
   structure.

ANTI-PATTERNS (never produce these)
- Inferring interest from duration alone
  (10-min showings can be hot; 40-min can
  be cold)
- Marking the showing complete in CRM
  before the agent has reviewed
- Auto-scheduling a follow-up without
  agent approval
- Voice memo errors propagating to CRM
  (always show the agent the extracted
  fields for review)
Open Claude →

iOS Shortcuts or a thin wrapper app handles the geofence + voice memo prompt. Claude extracts; the agent reviews.

built around this exact geofence-capture protocol

Letting the phone catch you is step one.
Reviewing the 5 fields is step two.

Lumi is the app that runs this workflow for you. You speak after a showing — Lumi captures the soft signals. You forward an email — Lumi updates the constraints. You open the app at 8am — the brief is already there, ready to feed Claude.

  • Voice → structured CRM, automatically
  • No forms. No data entry. No copy-paste.
  • Free for agents in EU · LatAm · MENA
9:41

Lumi · Wednesday

Good morning, Niki.

Two showings · three leads need a nudge.

Clara Ruiz
Tomorrow 11am showing at Passeig de Gràcia 84 with Clara Ruiz. She wants to bring her partner.
Got it — creating the showing.
Suggested event · 92%

Showing · Passeig de Gràcia 84

Thu · 11:00–11:45Gràcia
What’s the HOA for Apt 4?
€210 per month, covers elevator, concierge, and rooftop.DOC 12
Ask Lumi or speak…
Calendar
Todos
Lumi
Clients
Settings

A real-estate adaptation of the ambient-capture pattern from sales-enablement and field-ops — capture at the moment data is freshest, not 4 hours later. Our slice: iOS geofence + voice memo at the showing exit, with the structured brief in the CRM before the agent reaches the car.

More guides like this on @lumi.estate. Follow if any of this was useful — it's how we know to keep writing.