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.

10-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

A stranger becomes
a dossier. In 30 seconds.

Most agents stop at the name. The form fires, the CRM creates a card, the card sits empty until the agent finds 10 minutes to Google the person and write notes. The window of attention has already closed by then. The agents who triple their reply rate are the ones whose CRM cards are pre-briefed before they pick up the phone.

10-min readUpdated April 2026Pack 01 of 30 · @lumi.estate
marina_costa_brief.txt — preview
Marina Costa, Staff Engineer at Neuralink
(Porto, 4y), relocated from Lisbon 6 weeks ago.
House-hunting in Foz do Douro: 3-bed + garden,
likely a family-stage move. Flag: commute and
school district will outweigh sea view.
→ "Marina — three Foz 3-beds came on this week
   within 15 min of Tech Park. Saturday slots?"

Four sentences. One opening message. Total time from form submission to draft-on-screen: 47 seconds.

The gap between a form and a brief.

A real-estate form is the thinnest possible signal. Name. Email. Maybe a phone number, a budget range checkbox, and a free-text field that says “Looking at houses in Foz”. From this, an agent is supposed to figure out who this person is, why they're moving, what they actually want, and how to open the conversation in a way that doesn't sound like every other agent in the city.

The agents who do this poorly send a generic “Hi Marina, thanks for your interest! Here are five listings.” That message has a 6-9% reply rate in any market we've measured. The agents who do it well — the ones who triple that number — have a brief on screen before they hit send. They know Marina works at Neuralink, that she moved 6 weeks ago, that her commute will pin her to a 15-minute radius around Porto Tech Park, that she's been tweeting about schools.

They didn't research that themselves. The stack did. The brief arrived in their CRM before the lead's email had finished syncing. The first message went out within two minutes of the form submission, in their voice, with one specific detail that proved they'd done the work — even though they hadn't.

“A form is the thinnest signal in your pipeline. A brief is the thickest. Most agents close the gap manually — at the cost of every lead they don't have time to research.”

the protocol

Four steps. Sixty seconds.

Each step is a constraint that protects the speed-to-brief promise. Skip any one and the brief either arrives too late, contains too much, or doesn't end in a draft message.

  1. 01

    The form fires the stack — within 60 seconds.

    Zillow form, IG DM, WhatsApp inbound, website form — whichever channel the lead arrives on, the enrichment stack runs synchronously. The agent should never see a CRM card without a brief. If the brief takes 5 minutes to arrive, the agent has already moved on to a different task. Fast feels personal; slow feels automated.

  2. 02

    The enrichment finds 8 fields, not 50.

    Most enrichment APIs return everything they have — employer, alma mater, every job since 2008, every public mention. Agents drown in this. The whole game is choosing which 8 fields actually move a real-estate decision: commute, life event, family stage, channel preference, mover history, neighbourhood pattern, budget signal, urgency signal. The other 42 fields are noise.

  3. 03

    Claude writes the 4 sentences — not 8, not 2.

    Four is the sweet spot: enough room to compress a person into a brief, tight enough that the agent can read it in 8 seconds before picking up the phone. Two sentences feels thin and AI-generated. Eight sentences becomes a wall the agent will scroll past. The 4-sentence shape is the contract between the model and the agent.

  4. 04

    The fourth sentence is a draft message, not a summary.

    This is the rule that flips the whole workflow from "data assistant" to "opening assistant". The brief is not just for the agent's understanding — it's for the agent's first move. By the time the brief exists, the opening line already exists. The agent reads four sentences, taps approve on the draft, and the conversation is open within 90 seconds of the form submission.

the eight fields

What the stack looks for. What it ignores.

Public enrichment APIs return 50+ fields. Agents need 8. These eight — and the discipline to drop the other 42 — are what turns enrichment into intelligence.

field
why it moves the deal
current_city + tenure
Where they live now and how long they've lived there
Long tenure = local network = referral surface. Short tenure = recent mover = possibly moving again.
employer + role + tenure
Job, seniority, years in role
Income proxy, commute anchor, life-stage signal. Senior + new = likely to upgrade housing within 12 months.
prior_address
Last known address before this one
Distance moved tells you about the move's nature. Cross-country = job; cross-neighbourhood = lifestyle/family.
life_event_signals
Marriage, baby, divorce, parent move-in (public records + social)
The single highest-signal predictor of a real-estate decision in the next 6-18 months.
commute_target
Where they commute to (employer address + transit pattern)
Hard constraint. A 30-min commute and a 60-min commute are different searches entirely.
channel_preference
Email vs WhatsApp vs SMS vs IG — which they actually respond on
Inferred from how they submitted the form and their public-channel activity. Wrong channel = no reply.
neighbourhood_signal
Posts, follows, search history (consented) about specific areas
If they've been tweeting about Foz do Douro for 6 months, that's a higher-confidence target than the form's checkbox.
budget_signal
Implied range from employer + role + saved listings + life event
Most agents avoid asking budget early; the stack triangulates a range so the first listings sent are calibrated.
the stack

Four tiers. Trigger to delivery.

The stack is tool-agnostic — swap any one tier without rebuilding the others. The contract is the brief format, not the vendors.

Tier 1 · trigger

Zillow form / IG DM / WhatsApp inbound / website

Raw payload: name, email, phone, free-text message

Tier 2 · enrichment

Clay / Apollo / Crustdata / People Data Labs

Employer, role, tenure, prior addresses, public socials

Tier 3 · synthesis

Claude (Haiku for speed, Sonnet for nuance)

4-sentence brief + opening message draft

Tier 4 · delivery

CRM custom field / Lumi inbox / Slack / WhatsApp Business

Brief surfaces in the agent's interface within 60 seconds

anti-patterns

Three briefs the model defaults to without rules.

When the system prompt is loose, the synthesis tier produces one of these three failure modes. Each one looks like a brief but doesn't act like one.

the bot

Hi Marina! I noticed you submitted a form on our website looking at houses in Foz do Douro. As a Neuralink employee, you might want to know about our exclusive listings near Porto Tech Park. Let me know if you're interested!

Mentions the form ("submitted"), name-drops the employer in a vaguely creepy way, ends with the dead "let me know if you're interested". This is what agents send when they don't have a brief — the prompt-without-context version of the same input.

the data dump

Marina Costa, 4-year tenured Staff Engineer at Neuralink. Moved from Lisbon to Porto 6 weeks ago. Previously lived at Avenida da Liberdade 234. Married in 2024. Currently looking at 3-bed houses in Foz do Douro with a garden. Commute target: Porto Tech Park (15 min driving). Active on Twitter about Porto schools.

Reads like a CSV. The agent now has to do the synthesis themselves — what's the angle? what should I send? The whole point of the brief is to do the synthesis IN the brief.

the over-reach

Marina is a Staff Engineer at Neuralink who recently relocated to Porto and is likely planning to have children based on her Twitter activity. We have several premium listings she would love.

Inferring "planning to have children" from public Twitter posts is creepy and probably wrong — soft-signal pattern matching is a tool, not a fact. The brief must surface signals as flags for the agent to weigh, not promote them to certainties.

copy · paste

The enrichment payload.

This is the JSON the synthesis tier receives. Notice that some fields are empty — the prompt handles that gracefully by skipping them rather than naming the gap.

enrichment_payload.yaml
# ── enrichment payload — input to the prompt ────────
name:            "Marina Costa"
email:           "marina.costa@neuralink.com"
phone:           "+351 91 234 5678"
employer:        "Neuralink"
role:            "Staff Engineer · Implants team"
tenure_yrs:      4
linkedin_signal: "moved Lisbon → Porto 6 weeks ago"
prior_address:   "Avenida da Liberdade 234, Lisbon"
commute_target:  "Porto Tech Park (15 min driving)"
public_record:   "married 2024, no children listed"
soft_signal:     "active on Twitter about Porto schools"
form_message:    "Looking at houses in Foz do Douro,
                  3-bed minimum, garden if possible"
the prompt that writes it

What to feed Claude.

The system prompt that turns the enrichment payload into the 4-sentence brief. Tested against Haiku and Sonnet — Haiku is fast enough to fit inside the 60-second SLA on first pass.

dossier_system_prompt.md
You are a senior real-estate agent's lead-enrichment
analyst. Your job is to turn a raw enrichment payload
into a 4-sentence brief that fits in a CRM card.

INPUT
You receive a JSON object with whatever fields the
enrichment stack found: name, email, phone, employer,
role, tenure, commute estimate, prior addresses,
LinkedIn snippets, social signals, public records.
Some fields will be empty. That is normal.

OUTPUT
Write EXACTLY 4 sentences. Plain text. No bullets,
no headings, no emoji.

  1. WHO they are — name, current city, role + tenure,
     anything that grounds them as a person.
  2. WHY they're moving (or might be) — life event,
     job change, family stage, prior-address pattern.
  3. WHAT to flag — one constraint or preference the
     stack revealed (school district, commute, parking,
     pet policy, accessibility, prior-build pattern).
  4. WHERE to start — the single line the agent should
     send first, in their voice, no more than 12 words.

RULES (non-negotiable)
1. If a field is empty, skip it gracefully — never
   say "not found", "unclear", "to be confirmed".
2. Never invent a fact. If the stack didn't surface
   it, don't include it.
3. The fourth sentence is a draft message, not a
   summary. It must be sendable as-is over WhatsApp
   or SMS.
4. Voice: clinical-warm. Senior agent, not a chatbot.
5. No "this lead", "this prospect", "client". Use
   their first name from sentence 1 onward.

ANTI-PATTERNS (never produce these)
- "Based on the data provided…"
- "It appears that…"
- "Hi {{first_name}}, I noticed you submitted a form…"
- Any sentence longer than 22 words
- Any sentence that restates the input verbatim

The agent should read the 4 sentences in 8 seconds
and know exactly how to open the conversation.
Open Claude →

Copy the system prompt above into a new Claude chat as a system message, then paste the enrichment payload as your first user message.

payload in · brief out

What Claude returns.

Run the payload above through the prompt above. This is the first-pass output — no editing.

output · 4-sentence brief

Marina Costa is a Staff Engineer at Neuralink (Porto, 4 years tenured) who relocated from Lisbon 6 weeks ago — likely a commute optimisation around the Porto Tech Park. She's house-hunting in Foz do Douro for a 3-bed with a garden, and her Twitter activity about Porto schools suggests she's planning to start a family within the next 12-18 months. Flag: she'll prioritise commute-time and walkability to a good kindergarten over square metres or sea-view. Open with: "Marina — three Foz do Douro 3-beds came on this week within 15 min of the Tech Park. Send Saturday slots?"

fields used
employer · prior_address · life_event · commute
flag surfaced
commute > sea-view priority
next step
Saturday slots — yes/no question
about the number

Where the 3× comes from.

The 3× reply-rate figure is a working benchmark across ~30 agents in EU and LatAm running a brief-driven first-touch versus a generic first-touch on otherwise comparable lead pools. Generic first-touch (“Hi Marina, thanks for your interest”) sits in the 6-9% reply range. Brief-driven first-touch — where the opener references one specific detail from the dossier — lands at 22-31%. The lift is usually 3-3.5×.

The honest caveat: the 3× is on inbound form leads where the enrichment stack actually finds something. For leads with thin public footprint (~15-20% of inbound, depending on market and channel), the stack returns mostly empty fields and the brief regresses toward the generic baseline. The protocol still helps in those cases — the agent at least knows the stack tried — but the lift is closer to 1.4-1.8× there. Average across the full inbound pool: 2.7-3.1×.

about the data

Public signal, not surveillance.

Every field in the enrichment tier comes from data the lead has already made public — LinkedIn employment, public Twitter posts, public records, the form they themselves submitted. The stack is a speed-up of work the agent could (and used to) do manually with 10 minutes of Googling. It is not a surveillance product.

Two boundaries the protocol will not cross: it does not ingest private inboxes, private DMs, or paid people-search dossiers sourcing from breached data. And it does not surface inferred attributes that feel invasive when stated back — “likely planning children based on Twitter” is the kind of inference the brief either omits or downgrades to a hedge (“may be family-stage”). The agent's credibility with the lead depends on the conversation feeling thoughtful, not creepy.

built around this exact dossier protocol

Building the brief is step one.
Trusting the 60-second SLA 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 lead-enrichment-as-synthesis thesis (Clay-style data stacks compressed by an LLM into a 4-line brief). Our slice: closing the 60-second gap between form submission and the agent's first message.

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