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
Lumi · Wednesday
Good morning, Niki.
Two showings · three leads need a nudge.
Showing · Passeig de Gràcia 84
Your past clients have
the leads. You never asked.
One past client's LinkedIn 2nd-degree network contains, on average, 12 people who will make a real-estate decision in the next 12 months. Most of them, the past client doesn't even think to mention. The agents who systematically mine these networks — and ask in the right way, at the right time — produce two-thirds of their pipeline this way.
Past client: André Costa Window: April 2026 (post-tax-return) Top ask: Beatriz Sousa · Director · Coimbra signal: first child announced 2026-04 Past client's message (drafted, awaiting send): "Beatriz — congrats on the news! If you ever want a no-pressure chat with someone good in Coimbra real estate, my agent A. is great. Send if useful."
One ask. One past client. One named connection. The 67% intro rate sits on the back of every line of this discipline.
One past client = twelve warm leads.
The funnel from a past client's network to a closed transaction is steeper than it looks. Each step has a measurable conversion rate; the whole system stands or falls on filtering aggressively at step one.
Five rules. One ask per quarter.
The discipline of the graph mining. Skip any one and the past client either ignores the request or stops replying altogether.
- 01
Run the graph against the right week of the year.
There are 6 windows in a calendar year that produce the highest density of life-event signals: post-summer (Sept), pre-school-year (Aug), new-year reset (mid-Jan), tax-return season (Apr), pre-summer-move (May), Christmas-holiday consolidation (Dec). Run the graph mining pipeline once per window, not continuously. Continuous noise; window-aligned signal.
- 02
Filter aggressively. The signal is in 12-18, not 200.
Past client has 200+ 2nd-degree connections in the same metro. The graph mining surfaces 12-18 with a 90-day life-event signal. The other 180 are noise — running the ask against them dilutes the past client's relationship with you and looks like you're farming them. Discipline cuts the list to one digit before the past client ever sees it.
- 03
The past client sends. The agent does not.
Crucial: the agent never reaches out to a 2nd-degree connection cold. The past client sends the ask in their voice, with the agent as a name they recommend. The whole warmth comes from the past client's standing — bypass that and it's just another cold message with extra steps.
- 04
One ask per past client per window.
Maximum once per quarter. Each ask is for ONE specific connection (the highest-ranked one in that window). Past clients tolerate one specific intro request without it feeling like a chore; six requests over six months trains them to stop replying. The discipline is what protects the relationship.
- 05
The ask script is for a chat, never a transaction.
The past client's message offers a coffee, a 30-minute introductory chat, a contact for the connection's back pocket. Never a listing, never a CMA, never a buyer's-agency agreement. The transaction is downstream of the conversation; the conversation is downstream of the warm intro. Skip a step and the whole sequence collapses.
Three asks that cost the relationship.
The classifier produces these without strict prompt rules. Each one has been sent by a real past client and each one ended the agent's warm-intro pipeline with that person for at least 18 months.
“Hugo, my agent A. is doing a special promotion this month for newlyweds. If you and your partner are looking at houses, she has 5 listings in your area. Want me to forward her the intro?”
Sounds like the past client is being paid to refer (which they're not). Lists a promotion. Mentions specific listings. Every line erodes the warmth that justified the ask in the first place.
“Hugo, I noticed your wedding photos and wanted to congratulate you. As newlyweds, you might be thinking about your first home together — A. is the agent I worked with and she's amazing.”
"I noticed" reveals the watching. Names the inference ("newlyweds → first home"). Reads like a templated friend-of-friend pitch. The past client should never use this version — and our prompt actively rejects it.
“Hi everyone, I closed on my home with A. last year and she was incredible! If anyone is thinking of buying or selling, definitely reach out — happy to make intros!”
Mass-broadcast referral asks have a ~3% reply rate. Specific named-person asks have a ~67% reply rate. The 22× difference is the whole point of the protocol — running it as a broadcast is throwing away its main lift.
The graph snapshot.
What the prompt receives per past client per window. The 2nd-degree connections list is pre-filtered to same-metro + 90-day signal — the prompt does ranking and ask-line drafting only.
# ── referral graph snapshot ──────────────────────
past_client:
name: "André Costa"
closed_with_us: 2024-09 (€480k Lapa 2-bed)
network_size: 147 LinkedIn 2nd-degree
window_days: 90
connections (top 12 by raw signal):
- name: "Marina Costa"
employer: "Neuralink (Porto)"
role: "Staff Engineer · Implants"
tenure: "4 years"
signal: "moved Lisbon → Porto 6 weeks ago"
rank_raw: high
- name: "Hugo Almeida"
employer: "Block.one (Lisbon)"
role: "Head of Legal"
tenure: "2 years"
signal: "married 2026-02, public photo"
rank_raw: medium
- name: "Beatriz Sousa"
employer: "Critical Software (Coimbra)"
role: "Director of Engineering"
tenure: "8 years"
signal: "first child announced 2026-04"
rank_raw: high
- name: "Tiago Reis"
employer: "Volkswagen Autoeuropa"
role: "Plant Manager"
tenure: "12 years"
signal: "no recent change"
rank_raw: low
... 8 more connections truncated
What to feed Claude.
The prompt does ranking by life-event signal weight and drafts each ask in the past client's voice (not the agent's). Voice samples from the past client's prior messages are required.
You are a senior real-estate agent's
referral-graph analyst.
INPUT
You receive: one past client's name and the
public LinkedIn profiles of their 2nd-degree
connections (up to 200), each with: current
employer, role, tenure, location, and any
public life-event signal from the last 90
days (job change, marriage, baby, parent
move-in, retirement, location change).
OUTPUT
Rank the top 8 connections by likelihood
of needing a real-estate agent in the next
12 months. For each, output:
name + employer + role
proximity_signal: <which life event or
career signal raised
them above the noise
floor>
ask_line: <the single sentence the past
client should send when
reaching out — references the
mutual connection's name and
a low-pressure offer>
RULES (non-negotiable)
1. Job change in last 90 days outranks
every other signal — relocations cluster
here.
2. Marriage / new child / parent move-in
each move someone up by 2 ranks.
3. The ask_line is in the past client's
voice, not the agent's. The past
client is the one sending the message.
4. Never quote the connection's public posts
directly — reference patterns, not
sentences.
5. The ask is for a CONVERSATION, never for
a transaction. "Coffee", "intro", "happy
to help if useful" — never "find them a
home", never "list with us".
ANTI-PATTERNS (never produce these)
- Anything starting with "I noticed you…"
- "I'd love to introduce you to my agent"
(too direct, kills the warmth)
- Mentioning specific listings or prices
- Quoting the connection's public posts
- Any sentence longer than 22 words
A 2nd-degree connection should read the ask
in 5 seconds and feel like a normal note from
the past client — not a referral-broker pitch.Paste the system prompt as a Claude system message, then feed each past client's graph snapshot per window.
What Claude returns.
Top 3 (of 8 ranked) asks per past client per quarter. The agent reviews; the past client sends — never the other way around.
1. Beatriz Sousa · Critical Software · Director Engineering
proximity_signal: first child announced — typical
space-upgrade window opens 6-12 months
ask_line: "Beatriz — congrats on the news! If you ever
want a no-pressure chat with someone good in
Coimbra real estate, my agent A. is great.
Send if useful."
2. Marina Costa · Neuralink · Staff Engineer
proximity_signal: relocated 6 weeks ago — settling-in
window where rentals roll into purchases
ask_line: "Marina — saw you're in Porto now! If you're
renting and thinking longer-term, A. helped
me hugely in Lisbon. Want her contact?"
3. Hugo Almeida · Block.one · Head of Legal
proximity_signal: recently married — common 6-18 month
family-housing window
ask_line: "Hugo, congrats again on the wedding. If you
two ever start looking at houses, my agent
A. is the one I'd recommend. Happy to intro."Mining the graph is step one.
Trusting the past client to send 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
Lumi · Wednesday
Good morning, Niki.
Two showings · three leads need a nudge.
Showing · Passeig de Gràcia 84
Pipeline
Active
8
Warm
4
Cold
2
Clara Ruiz
Active€1.8M · 3BR
Passeig de Gràcia showing · 11:30
Andreas Moreno
Active€2.4M · 4BR
Send comps by 18:00
Dimitri Schneider
Warm€900K · 2BR
Contract review today
Silent 3d · last 3 days ago
Sarah Mitchell
Cold€1.2M · 3BR
Draft re-engagement
Silent 9d · last 9 days ago
A real-estate adaptation of the hidden-network thesis from sales — past customers' 2nd-degree networks at life-event moments outconvert any cold list. Our slice: one specific named ask per past client per quarter, with the past client as the sender.
More guides like this on @lumi.estate. Follow if any of this was useful — it's how we know to keep writing.