How to run the realtor workflow without Zapier (Phone-First Workflow)
Most realtors don't need Zapier. They need one prompt and a Saturday.
The 12-deal sub-threshold problem
A solo at Compass Hendersonville. Twelve sides last year — close to the NAR median of 10. Sphere-driven pipeline, no Zillow Flex, no paid lead-gen budget. No transaction coordinator.
She reads a "30 must-have real estate AI workflows" listicle and feels behind. She buys Zapier Pro and wires four Zaps her first weekend. By month four, 30% are dark. A Gmail token expired. Compass One pushed an update. She stops trusting any of it. Now she's checking everything by hand, plus paying $30/mo for the privilege.
She shouldn't have bought the stack. She's below all three thresholds the pillar lays out — 30 sides a year, $500/mo in paid leads, or a second human on the file. Below any one, the wrapper-stack is debug debt with marketing copy on top.
What she needs is small, judgment-light, and verifiable every step. A model on her phone, a Context Card pinned at the top of the thread, and a Saturday to learn the pattern.
What you'll learn
Sarah's Saturday open-house pre-prep at Compass Hendersonville. 90 minutes, MLS upload to ready-to-send drafts. One LLM agent — Claude or ChatGPT — in your hand. $20/mo. Four steps. Re-run weekly.
Step 1 — Pin a Context Card at the top of your Claude or ChatGPT thread
The Context Card is the difference between the model writing like you and writing like a SaaS chatbot. Four sections — role, voice, do-not-say, market knowledge. You write it once and paste it at the top of every new thread.
Open Claude (or ChatGPT Plus). Start a new conversation. Paste this as your first message:
Role: I'm a Hendersonville buyer's agent at Compass.
Most clients are buying $300K-$750K, primary residences,
mix of waterfront on Old Hickory Lake and inventory in
Cool Springs and Williamson County.
Voice: warm, direct, no jargon. Real-estate native always —
listing, comp, MLS, days-on-market, pre-list. Texts read
like a friend who happens to be your agent.
Do not say: lovely, stunning, nestled, charming, dream home,
unlock, leverage, supercharge, seamlessly. No exclamation marks.
Local knowledge: Old Hickory Lake church-lot parking fills
up by 10:30 on Saturdays. Williamson County school
districts matter to relocators. TREC requires me to ask
about working with another agent at first contact.
Acknowledge this card. From here on, draft everything in
this voice unless I tell you otherwise.
Hit send. The model echoes back the card. Every draft in the thread now inherits the rules. Save the thread. Bookmark it. Next time you open Claude on your phone, reopen this thread instead of starting fresh.
That's the pin. Five minutes, one time.
The four-layer structure isn't accidental. It maps to what actually drifts when a model has no grounding — voice, vocabulary, banned phrases, and the local detail nobody outside your market knows. The card makes those mechanical instead of vibes-based.
Step 2 — At the start of any task, paste raw context
This is the part most agents fumble. They summarize the lead instead of pasting it.
Don't summarize. Paste raw — the actual message, the calendar block, the lead form, the MLS print-out. The model is faster and more accurate reading raw data than backfilling from a summary you wrote in 30 seconds.
For a Saturday open-house run, raw context looks like:
- Three Zillow lead emails from Friday, copy-pasted whole
- The MLS sheet for 142 Lakeshore — bullets are fine
- Last 90 days of Old Hickory Lake comps from your Compass One report
- The Wilson family's iMessage thread — the part where they confirmed 1 PM Saturday at a different listing
You're not curating. You're dumping. The model sorts.
This is what the O'Reilly What We Learned from a Year of Building with LLMs group — Yan, Husain, Bischof, Frye, Liu, Shankar — call out as the pattern that moves agent reliability: small tasks with clear objectives, fed dense raw context. AlphaCodium went from 19% to 44% on CodeContests by feeding the model more raw input and letting it decompose the work itself. Same shape works on a phone for a Saturday open house.
The BiggerPockets ChatGPT for Real Estate guide lands the same point at the practitioner level — the agents getting real lift paste full lead emails and full MLS sheets, not a sentence and a hope.
If the data has client info you wouldn't want logged, scrub names and numbers before pasting. The model doesn't need them to draft a reply — it needs the property, the ask, and the timing.
Step 3 — Ask the model for the deliverable, then verify
Now the prompt itself. Short. Specific. Asks for a deliverable, not a chat.
Sarah's actual Saturday morning prompt, after the Context Card is pinned and the raw context is pasted:
Open house prep — 142 Lakeshore, 11 AM today.
I have 90 minutes. Generate four things in this order:
1. One-page market-update flier for visitors. Old Hickory
Lake submarket, last 90 days. Median sale price,
days-on-market, active inventory. One sentence on
what it means for a buyer right now. Tone: neighbor
at the kitchen counter.
2. Three follow-up texts to Friday's Zillow leads:
- Dan, asked about waterfront under $750K
- Priya, asked about a Hendersonville condo
- The Reeds, asked about 88 Cedarwood
Two specific time windows each, pre-approval ask,
TREC working-with-another-agent ask, my Context
Card voice.
3. Drive-to address pinned: 142 Lakeshore, plus the
church-lot parking note (front fills up by 10:30).
4. Showing brief for the Wilsons at 1 PM in Cool Springs —
different family from the open house. Three questions
to ask in person, two comps to mention, soft pre-
approval ask.
Also flag any weak signals — anything in the Friday leads
that hints at deeper interest, urgency, or a referral
opportunity I might miss skimming fast.
Output: ready to paste, in that order.
What comes back: a flier, three texts, an address with parking note, a showing brief, and a flagged weak signal — like Priya mentioning her cousin is moving to town in the spring. A paid lead form would never capture that. A model reading the raw email does.
The verify-then-send pattern is what makes this safe. You read every output before it goes anywhere. Simon Willison's weblog has hammered one point through 2025 — verification is the expensive part. The model drafts fast. The cost is making sure the draft is right. On a phone, in your hand, it's right there. Scan, fix a phrase, send.
Karpathy's "vibe coding" framing on X makes the same point about throwaway code — fine when you eyeball every line, dangerous when output goes straight to a client without a human reading it. The Phone-First Workflow keeps you in the loop on every send. That's the feature, not the bug.
Step 4 — Send. Bookmark the prompt. Re-run weekly with fresh context
The cheapest step and the most overlooked.
After Sarah edits the three texts and sends them, she does three things:
- Saves the conversation in her Claude app (it stays in thread history by default — confirm it's not set to auto-delete)
- Pins the prompt template to her Notes app under "Saturday open house prep"
- Sets a Friday-night calendar reminder to re-run it next week with new leads
Now the workflow compounds. Next week she opens the saved thread, swaps the raw context for the new week's leads and MLS sheet, runs the same prompt, gets the same shape of output. Voice stays consistent — the Context Card is still pinned at the top. Numbers update themselves because she's pasting fresh comps. The prompt is the program. The model is the runtime.
What the result looks like
By 8:30 AM Saturday, Sarah's prep is done:
- Three Friday Zillow leads have a reply, well inside the LRM 5-minute window from when she picked up her phone
- A market-update flier ready to print at Compass at 9 AM
- 142 Lakeshore pinned in Maps with the parking note
- A Wilson-family showing brief in Notes for the 1 PM drive
- One flagged weak signal she'd have missed — Priya's cousin moving to town
Total model time: 7-12 minutes per session. Total stack: $20/mo Claude Pro (or ChatGPT Plus), the iPhone she already owned, a Context Card she wrote once. Zero re-auth. Zero OAuth tokens. Zero debug surface.
Compare a Zapier setup at the same volume — $30/mo for a stack that needs weekly walk-throughs to confirm it's alive, plus the four-Zap setup tax of a Saturday she'd never get back.
Why this works
Three things have to be true for the Phone-First Workflow to beat the wrapper stack at sub-threshold volume.
One — small tasks with clear objectives. From Yan, Husain et al. in What We Learned from a Year of Building with LLMs — Part I, the pattern that works at the model layer is small, focused, decomposed tasks. "Generate three texts and a flier" is small. "Run my whole follow-up business autonomously" is not. The Phone-First prompt sits where models are reliable.
Two — verification is in your hand. Simon Willison's weblog made this case through 2025 — draft speed is free, verification cost compounds. A Zap that runs unsupervised has no verification step. A model that drafts on your phone and waits for you to send has verification baked in.
Three — English is the program. Karpathy's Software 3.0 framing on X lands the bow on it. The prompt is the program. The model is the runtime. No wrapper, no platform, no integration to maintain.
The wrapper stack is optimized for volume the median agent doesn't have. The Phone-First Workflow is optimized for the volume she actually does.
When this DOESN'T work
This page is for the agent below all three thresholds. Cross any one and the math flips.
- 30+ closed sides a year. NAR median is 10. Three times the median is when data drops in the seams between deals, and a real database earns its keep.
- $500+/mo in paid lead-gen. Roughly 30+ inbound leads a month from Zillow Flex, Realtor.com Connections, or Compass paid programs. Past that, the LRM 5-minute window can't be hit manually outside business hours.
- 2+ humans on the deal. A TC, ISA, or buyer's agent means the system of record can't be your text thread. They need shared visibility.
Cleared any one? Wire the three Zaps a 12-deal agent actually needs — lead intake, showing handoff, contract dates. 30 minutes of one-time setup.
For a head-to-head on which platform earns the spend, see Zapier vs Make.com vs the foundation-model approach.
If you're staffing up, paid-lead ratio is climbing, and the back-office is leaking — the question shifts from "do I need automation" to "how do I build the system around it." That's what The Listing Machine operationalizes — four-week beta cohort, AI-Enhanced Realtor credential, prompt stack tuned to your voice and your market.
Below the thresholds — don't buy the wrapper stack. Run the Saturday workflow. Bookmark the prompt. Re-run weekly. That's the job.
Sources
Primary data
Independent builder/operator creators
Practitioner reference
Last updated 2026-04-29.
Schema (HowTo, four steps)