For working agents who already use ChatGPT — and get generic answers back

Claude reads your market files.
Five jobs. Four populated markets included.

A folder of markdown files turns Claude into a residential-real-estate specialist for your city. Five jobs. About 20 minutes to set up. Free, MIT-licensed, no install, no API key, no monthly cost beyond your Claude subscription.

Recommended path · ~30 minutes

Start from a populated market.

Four real markets ship with v2. If yours is structurally similar to any of them, copy the region pack over and edit for your specifics. Faster than the 20-minute fill-in, and you see the format before you commit.

Recommended

Path A — start from a case study pack

Open case-studies/ and pick the market closest to yours by shape — diaspora-driven (Jerusalem), universal-no-foreign-hook (Novato), foreign-buyer-dominant resort (Khao Lak), or mainstream urban European (Lisbon). Copy its region folder over the empty one and edit. You'll be running real jobs in under 30 minutes.

Path B — start from scratch with six placeholder files — works too, and is documented step-by-step in the repo. Most agents pick A.

What it does

Five jobs. One folder. That's the whole thing.

Every new chat opens with a menu. Pick a job, hand it the inputs, get back a structured artifact you can defend to a client or a broker.

JobYou hand it...You get back
Write a listing Property specs + audience (e.g. "international investor") A finished listing tuned to that buyer pool. Multiple languages on request.
Match a buyer★ most-used job Buyer profile + 3–4 candidate listings Ranked fit with reasoning, friction points called out, which one to show first.
Price a property Property specs + target window Defensible price band (low / mid / high), comp set, adjustment logic.
Capture showing notesnew in v2 Date, buyer ID, your dictation transcript Structured note, friction tags, agent-facing read. Cited back by future jobs.
Look up a servicenew in v2 Service type + property context Your top pick from your own curated vendor list, with the reasoning. Never invents.
Four populated markets tested

Four populated markets. Real outputs.

Same architecture, four very different markets. Same five jobs, working outputs. Here's what comes back when you actually run a job.

Job 1 — Listing Jerusalem · international investor
Old City–adjacent 4-unit residential building → English/Hebrew listing pair
You ask
Listing job. 4-unit residential building, 5 minutes' walk from Jaffa Gate. ~340 m² total, fully occupied, current monthly rent ~$11,000 USD aggregate. Asking ~$3.4M USD. Audience: Chinese investor, English-speaking intermediary, yield-focused.
You get back
  • An English/investor-framed listing with yield calculation, location thesis, and exit-liquidity note.
  • Foreign-buyer tax considerations surfaced (Mas Rechisha, Mas Shevach) — not invented, drawn from the region pack.
  • A side-by-side Hebrew local-market listing.
  • Agent-facing prep notes for the showing — residency status, source-of-funds path, hold thesis, decision-pace expectations.
Job 2 — Buyer match Jerusalem · Modern Orthodox aliyah family
Three listings → ranked fit with reasoned friction
You ask
Matching job. Buyer: Family of 4 relocating from NYC (aliyah). Modern Orthodox; walking-distance shul + day school is critical. Budget $1.5M (stretch $1.65M). Three listings: A) Talbiya 4BR $1.7M walk-up; B) Baka 4BR $1.45M ground-floor garden; C) Old Katamon 3BR + office $1.55M elevator. Rank fit.
You get back
  • Ranked table — Listing C top pick, B at "Friction" (specific reason), A at "Blocker" (specific reason).
  • Detailed scoring on the top pick across Budget / Location / Type-size / Lifestyle / Friction.
  • Honest tradeoff section ("they asked for 4BR, this is 3+office").
  • A "don't dismiss yet" callout on Listing B with a real case for a second showing.
Job 3 — Pricing band Novato / Marin County, CA
4BR/3BA renovated single-story → defensible band with comp logic
You ask
Pricing job. 4BR/3BA in Novato (Pleasant Valley), 2,150 sqft, 0.18 acre lot, 1978 single-story, kitchen + primary bath remodeled 2021. Owner targeting on-market within 3 weeks.
You get back
  • Defensible band: Low $1,225K / Mid $1,275K / High $1,325K — with when-to-use guidance for each.
  • Comp-set table — 4 properties with sqft, sale price, $/sqft, sale date, adjustments to subject.
  • Adjustment logic with reasoning, drawn from the local market file you populated.
  • Market context (days-on-market, season, buyer-pool dependency) and a short caveat to verify before pricing.
Job 4 — Showing notes new in v2 Novato · empty-nester downsizers
Raw dictation → structured note with cross-job memory
You ask
Showing notes job. Showing was 2026-05-04 at 142 Bel Marin Keys Blvd. Buyer: the Hendersons (empty-nesters from Tiburon, $1.5–1.8M). Here's my dictation: [paste]
You get back
  • Structured note with property/buyer header, reaction summary, positive signals, friction (each line marked "from transcript" or "agent's read").
  • Comparisons, questions raised, logistics — separated cleanly.
  • Friction tags (kitchen-layout, hoa-cost, flood-risk) so future matching/pricing jobs cite it back.
  • Agent's read: "Linda's kitchen objection is the real obstacle, not the dock — lead the second showing with that."
Already populated

Four geographies. Four different proof points.

Each ships with a populated region pack you can copy and edit. The diversity is the point — the architecture has to work across very different market shapes, not just one.

Jerusalem, Israel

Multi-language (Hebrew + English), strong international diaspora flow, complex title types (Tabu vs. Khevra Meshakenet). Proves the multilingual + foreign-buyer wedge.

Novato / Marin County, CA

English-monolingual U.S. suburban market — proves the architecture works without any foreign-buyer dependency. The "could this work in my normal American suburb?" answer.

Khao Lak, Thailand

Coastal resort, foreign-buyer-dominant, leasehold/freehold structure, multi-language (English / German / Russian / Mandarin). Proves the resort/expat market shape.

Lisbon, Portugal new in v2

Mainstream urban European international-buyer market, post-Golden-Visa context, multilingual (Portuguese / English / French / Brazilian Portuguese), Roman-law title system.

Three things a generic chat session can't do here

Three places. Honest about the rest.

Most AI-for-agents marketing claims to beat ChatGPT at everything. This claims three things, and tells you where ChatGPT is fine.

1. Buyer–property matching with reasoned friction

ChatGPT keyword-matches ("4BR matches 4BR"). The copilot reasons about intensity of fit, surfaces what the buyer will compromise on for the top pick, and tells you when no listing in your candidate set actually works.

2. Price bands with citable comp logic

Not "list at $X" — a band, with the comp set you can defend on a listing presentation, with adjustment magnitudes pulled from your local market file.

3. Cross-job memory

Showing notes you capture today can be cited in a matching job next week. The friction tags carry forward. ChatGPT forgets every chat.

Where ChatGPT is fine

Short marketing blurbs, generic Q&A, idea brainstorming. The copilot doesn't compete on those — there are dedicated tools and ChatGPT itself does the job. Save the copilot for the work that needs structure.

Why it knows your market

Three files. Three jobs. That's the mechanism.

The copilot isn't a model trained on real estate. It's Claude, reading three layers of plain-text files in the right order. Your market file is the layer that makes the answers yours.

01

Specialist (universal)

Identity, rules, examples, frameworks. The "how this works" that's the same in every market. Already written, already loaded — you don't touch it.

02

Region pack (yours)

Six files describing your market: inventory, neighborhoods, regulations, contracts, glossary, services. Fill in once. Update when something changes.

03

Cross-job memory

Showing notes get tagged with friction terms. The next time you run a matching or pricing job for the same buyer or property, those tags are visible to the copilot.

Hard-bounded scope

What it won't do.

The copilot redirects out-of-scope requests on purpose. Five jobs. Anything else, it points you to the right tool.

Out of scope, by design

  • Won't give binding legal or tax advice. Surfaces what to ask your lawyer or accountant about.
  • Won't quote a single list price. Always a band, always with comps.
  • Won't draft client-facing emails (seller updates, buyer follow-ups). That's CRM territory.
  • Won't recommend vendors that aren't in your own services list. Never invents.
  • Won't write blog posts, social copy, or cold-outreach campaigns. Out of scope.
  • Won't try to do anything outside the five jobs.
How you use it

Three steps. About 20 minutes.

No code. No install. No API key. If you can use Claude in a web browser, you can use this.

01

Download the repo

On the GitHub page, click Code → Download ZIP. Unzip. You'll have a folder with specialist/, case-studies/, and docs/ inside.

If you've used GitHub before, git clone works too. Either way, no install — it's just files on your computer.

02

Fill in six files for your market

Open specialist/reference/region/. Six files with section headers and [PLACEHOLDER] markers: market, neighborhoods, regulations, contracts, glossary, services. Replace placeholders with your local data.

Or — much faster — copy a case study pack over and edit. Partial fills are fine; the copilot will tell you when it's missing something it needs.

03

Drop into a Claude Project and start a chat

Go to claude.ai. Create a new Project. Drag the specialist/ folder into Project knowledge. Start a chat.

The copilot greets you with the welcome menu — five jobs to choose from. Pick one. Hand it the inputs. You're working.

Ready to set up your market?

Free. MIT-licensed. The whole thing is a folder of markdown files. About 20 minutes from download to first real answer.

For developers, OSS folks, and judges

Under the hood: a folder of markdown.

The architecture is called Interpretable Context Methodology (ICM) — folders as architecture, files as the program. The structure is the documentation.

The whole thing, at a glance

your-market-realtor/
├── specialist/              # the universal part — same in every deployment
│   ├── welcome.md          # first-turn menu
│   ├── identity.md         # who the copilot is, what it does, what it won't do
│   ├── rules.md            # routing logic, file-loading discipline, "don't drift"
│   ├── examples.md         # worked examples of all five jobs
│   └── reference/
│       ├── region/          # YOURS — 6 files you fill in for your market
│       ├── frameworks/     # buyer-fit, pricing-comp, listing structure, etc.
│       ├── checklists/     # per-job preflight
│       └── templates/      # output shapes
│
├── case-studies/           # 4 populated region packs — start here
│   ├── jerusalem/
│   ├── novato-marin/
│   ├── khao-lak/
│   └── lisbon/
│
└── docs/                   # architectural spec, press kit, outreach package

No code. No install. No API. The routing in specialist/rules.md tells Claude which framework files to load for which job — so every request gets only the files it needs, not the whole folder.

For a deep technical walkthrough, see docs/v1-to-v2-spec.md. The pattern is portable — fork it for any narrow professional specialist (financial advisor, paralegal, clinical case manager) and the architecture transfers.