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

Every IG comment
is a lead. You're losing 90%.

Most agents post listings on Instagram and treat the comments as engagement metrics. But every “is the kitchen south-facing?” and every “how far to the train?” is a high-intent signal — usually higher than the cold leads agents pay for. The problem is reply speed: comment buyers go cold within an hour.

9-min readUpdated April 2026Pack 04 of 30 · @lumi.estate
auto_reply.dm — 47s after comment
@isabella_lx (commented 47s ago):
  "Is the kitchen south-facing? And how far
   to the train?"

draft (auto-sent):
  "South-facing — best light in winter. Train
   is 8 min walk, runs to Lisbon every 20.
   Saturday at 11 work for a viewing?"

Two sentences. Both her questions answered. One yes/no question. CRM lead created in parallel with full context.

The 90% you're losing.

Look at any agent's last 10 IG listing posts. Count the comments. Of those comments, how many got a DM reply within an hour? Across the agents we've audited, the answer is roughly 10-15%. The remaining 85-90% either got a reply hours later (effectively never) or got no reply at all.

The problem is not lack of will. The agent saw the comment, meant to reply, got pulled into a showing, came back, the comment was buried under three new posts. By the time they got around to it, the commenter had moved on, scrolled, forgotten. The comment-buyer window is roughly 60 minutes — measurable across thousands of cases — and almost no agent operates inside it.

The fix is not “reply faster”. The fix is removing the agent from the critical path on hot-tier comments entirely. The agent reviews and audits at the end of the day. The agent does not have to be in the loop for every “south-facing kitchen?” question — that's a question their classifier should answer in 90 seconds, in their voice, while they're showing a different property.

“The fix is not ‘reply faster’. The fix is removing the agent from the critical path on hot-tier comments entirely.”

the intent signals

What the classifier looks for.

Six patterns that separate a hot-tier buyer comment from a cold compliment from a spam bot. Each one gets weighted by the prompt; tier is decided on the strongest single signal.

signal
example
Explicit feature question
hot
"Is the kitchen south-facing?" — they have a feature checklist already.
Logistics question
hot
"How far to the train?" "What's the HOA?" — they're modelling the daily reality.
Life-event in bio
hot — defensive
Bio says "moving to Cascais", recent posts have moving boxes — high-confidence buyer.
Comparative question
warm
"How does this compare to the one in Estoril?" — they're shopping multiple properties.
Pricing without context
warm
"How much?" — they haven't seen the price; they want it. Reply with price + one specific feature.
Vague compliment
cold
"Beautiful!" "Love this!" — almost certainly not a buyer signal. Skip the auto-reply.
the protocol

Five steps. Ninety seconds.

The end-to-end flow: comment posted → CRM lead created with full brief → DM auto-sent in agent's voice. The agent never touches the critical path.

  1. 01

    ManyChat catches the comment within 30 seconds.

    ManyChat (or any IG comment-to-DM tool) fires when a comment hits the listing post. The webhook ships the comment text, commenter username, and post context to your stack within 30 seconds — well before the commenter has closed Instagram. Speed-of-reply is the single biggest factor in conversion at this stage; a reply within the same session feels like a person, a reply 2 hours later feels like an autoresponder.

  2. 02

    Public profile enrichment runs synchronously.

    Pull the commenter's public bio, follower count, recent post types, and (where allowed) public LinkedIn match. This adds 1-3 seconds to the latency budget but is what lets the classifier hit hot-tier confidence. Don't skip this step — without it, the classifier is reading a single comment in a vacuum and will misclassify ~30% of the time.

  3. 03

    Claude classifies + drafts in one call.

    Single LLM call: comment + profile + listing + agent voice samples in. JSON-shaped object with tier, signals, and draft reply out. Use Haiku for cost (~$0.0003/comment) and speed (~1.2s p50). The classifier and the drafter are the same prompt — splitting them adds latency without adding accuracy.

  4. 04

    Hot replies go out automatically. Warm queue for review.

    Hot-tier replies auto-send within the original session (target: <90s from comment to DM). Warm-tier drafts queue for the agent's review — these are the comments where the auto-reply could sound off, so the agent eyeballs and approves before sending. Cold and spam are silently logged and never replied to. The agent reviews the day's classifications in a 5-minute end-of-day pass.

  5. 05

    CRM card created on every hot/warm — even before reply.

    The moment the classifier returns hot or warm tier, a CRM lead is created with the commenter's profile, the listing context, the intent signals, and a 4-sentence brief (using the dossier protocol, see /prompt-dossier). By the time the agent sees the auto-reply went out, the lead is already in the pipeline with full context — no manual entry, ever.

anti-patterns

Three replies that kill the conversion.

The classifier produces these without enough voice samples or without the rules above. Each one is a real reply we've seen agents auto-send before tightening the prompt.

the bot

Hi! Thanks for your interest in this beautiful property. Please send me a DM and I'd love to chat about scheduling a viewing!

Doesn't answer the question. Doesn't reference the south-facing or the train. Reads like every IG-marketing autoresponder. The commenter sees this and assumes it's a bot — accurately.

the price-pusher

Yes! €825,000, fully renovated, 168m². The kitchen has been redone last year. When can you visit?

Leads with the price (the commenter didn't ask). Skips the question they did ask. Pushes a viewing before answering. Three failures in three sentences.

the over-eager

Great question! Yes, the kitchen is south-facing — perfect for plant lovers and morning coffee ☀️ The train is just 8 minutes walking distance, super convenient! Would you like to schedule a viewing this weekend? Saturday or Sunday work for you? I can also send more photos! Just let me know! 😊

Three exclamation marks and two emojis the agent doesn't use. Asks two questions in a row instead of one. Offers more photos before the commenter asks. Voice mismatch — the agent's actual replies are 1-2 sentences, no emoji.

copy · paste

The classifier input.

What the prompt receives on every comment. The voice samples are the single most important field — without them the reply will sound like a generic IG bot.

comment_input.yaml
# ── ig comment intent classification input ──────
comment_text:    "Is the kitchen south-facing?
                  And how far to the train?"
commenter:
  username:      "@isabella_lx"
  bio:           "Mum of two · Lisbon → Cascais soon"
  recent_posts:  ["beach photo", "kid's birthday",
                  "moving boxes"]
  followers:     412
  following:     387
listing_context:
  price:         "€825,000"
  neighbourhood: "Cascais — Praia do Tamariz"
  beds:          3
  bath:          2
  surface:       "168 m²"
  features:      ["sea-glimpse", "south-facing
                  kitchen", "8 min walk to
                  Cascais train", "garden
                  90m²", "garage 2 cars"]
agent_voice_samples:
  - "Yes, both bedrooms south-facing — best
     light in winter. Want a Saturday slot?"
  - "Garage fits two — one's a charging port.
     11am Saturday work for you?"
  - "Sea view from balcony, not from kitchen
     — sending the photo set. Still want it?"
the prompt that classifies + drafts

What to feed Claude.

One prompt does both jobs — tier classification and voice-matched reply drafting. Splitting them adds latency without accuracy gain.

dm_classifier_prompt.md
You are a senior real-estate agent's
IG-comment intent classifier and auto-replier.

INPUT
You receive: the comment text, commenter
username, public profile snippet (bio,
location, recent posts), the listing the
comment is on (price, neighbourhood, beds),
and the agent's voice samples (3-5 prior
DM replies the agent has actually sent).

OUTPUT
A JSON object with three keys:

  intent_tier: <one of:
                hot   — explicit interest in
                        seeing/buying the
                        property
                warm  — clarifying or info-
                        gathering question
                cold  — vague compliment, no
                        purchase signal
                spam  — bot, foreign-language
                        spam, irrelevant
               >
  intent_signals: <array of 1-3 specific
                   signals from the comment
                   or profile that drove the
                   tier classification>
  draft_reply:    <a 1-2 sentence DM reply,
                   in agent's voice, that
                   moves the conversation
                   forward IF tier is hot
                   or warm. Empty string for
                   cold/spam.>

RULES (non-negotiable)
1. Match the agent's voice samples — sentence
   length, punctuation rhythm, emoji policy
   (if the agent uses zero emoji, use zero).
2. NEVER mention the listing's price unless
   the commenter asked. Listing prices in
   first-touch DMs feel pushy.
3. Hot tier: end the reply with a 60-second
   yes/no question (a date, a viewing slot,
   "sending the floor plan?"). Never an
   open-ended "let me know".
4. Warm tier: answer the question, then drop
   ONE specific detail about the listing
   that wasn't in the comment (a feature
   the commenter would care about based on
   what they asked).
5. Cold/spam: return empty draft. The agent
   reviews and either ignores or sends a
   brief like-acknowledgement themselves.

ANTI-PATTERNS (never produce these)
- "Thanks for your interest in this property!"
- "I'd love to chat — DM me?"
- "Yes, this property is still available"
- Any sentence longer than 18 words
- Any emoji unless the agent uses them
- Generic "feel free to" / "let me know"

The commenter should read the reply in 4
seconds and feel like they got a useful
answer from a person — not a CRM autoresponder.
Open Claude →

Use Haiku for the latency budget. Pipe ManyChat webhook into the classifier; auto-send hot tier, queue warm for review.

comment in · reply out

What Claude returns.

JSON for tooling, plain prose for the agent's review queue. Tier hot → auto-send. Tier warm → queue. Tier cold/spam → silent log.

output · classifier json
{
  "intent_tier": "hot",
  "intent_signals": [
    "explicit feature questions (south-facing, transit)",
    "bio signals 'moving' — life-event match",
    "recent posts include moving boxes"
  ],
  "draft_reply":
    "South-facing — best light in winter. Train
     is 8 min walk, runs to Lisbon every 20.
     Saturday at 11 work for a viewing?"
}
built around this exact comment-to-CRM funnel

Replying in 90 seconds is step one.
Trusting the classifier 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 keyword-DM funnel pattern from ManyChat power-users in e-commerce. Our slice: IG listing-comment intent and the sub-90-second reply window where the conversion lift compounds.

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