off-market

---

name: off-market description: Use when the user wants to find off-market house-buying candidates — homes that aren't currently listed but where the owner may be motivated to sell (distressed, inherited, long-vacant, sheriff sale, tax delinquent). Two stages: /off-market discover <county> ranks candidates from free public data; /off-market outreach <candidates.csv> drafts personalized letters in the user's voice. v1 supports Allegheny County (Pittsburgh) via WPRDC; additional counties via plug-in adapters. Free data only — no PropStream / ATTOM / BatchLeads dependency. user-invocable: true

Off-Market House-Buying Skill

Personal toolkit for finding houses that aren't for sale yet but might be.

Two stages

1. Discover — /off-market discover <county> [--criteria criteria.yaml] Pulls free public data (parcels, sales, tax delinquency, sheriff sales), filters out actively-listed homes, scores each remaining parcel by sell-propensity, outputs a ranked CSV.

2. Outreach — /off-market outreach <candidates.csv> Walks top-N candidates with the user, drafts personalized letters in the user's voice using voice/profile.md.

Hard constraints

• Free data only.
• Never score on any federal FHA protected class (race, color, religion, national origin, sex, disability, familial status) — directly or by proxy. We also forbid age as a stricter internal policy. See references/fair-housing.md.
• No divorce / bankruptcy filings even though legal — ethically inappropriate for owner-occupier outreach.

Files

• scripts/discover.py — Stage 1 orchestrator.
• scripts/outreach.py — Stage 2 orchestrator.
• scripts/county_adapters/ — one file per county. v1: allegheny_pa.py.
• scripts/signals/ — one file per propensity signal.
• scripts/propensity.py — weighted scoring with fair-housing enforcement.
• references/ — propensity rubric, fair-housing rules, adapter checklist, letter craft.
• voice/profile.md — user's saved voice profile.
• examples/criteria.yaml — template buying criteria.

Known limitations (v1 — Allegheny County)

• Sheriff-sale data is stale. The WPRDC sheriff-sales dataset stopped updating in late 2022; "Current Bid List" maxes out at SaleDate=2022-12-05. The signal still fires (40 points) for historical auctions, but won't catch currently-distressed owners. To fix when actively using the tool: swap the WPRDC fetcher in scripts/county_adapters/allegheny_sheriff.py for a live scrape of https://www.alleghenycountypa.gov/services/sheriff/real-estate-sales/. Adapter architecture supports this without touching the orchestrator.
• Owner name is a code, not a name. WPRDC publishes OWNERDESC (REGULAR / CORPORATION) rather than actual owner names. The mailing address IS real, so letters reach the right house — they just can't address the recipient by name. Letter drafting (Phase 7) opens with a generic salutation. Workaround for higher-value targets: look up the actual name on Allegheny's Real Estate Portal manually before mailing.

How to use

# 1) Set up criteria — copy a starter template and edit it
cp off-market/examples/allegheny_pa.yaml my-criteria.yaml
$EDITOR my-criteria.yaml

# 2) Discover candidates (writes to ./runs/<date>-<county>/)
python3.11 off-market/scripts/discover.py allegheny_pa --criteria my-criteria.yaml

# 3) Draft letters for top 20 candidates
python3.11 off-market/scripts/outreach.py runs/2026-06-02-allegheny_pa/candidates.csv --top 20

Output structure per run:

runs/2026-06-02-allegheny_pa/
├── candidates.csv       # ranked candidates with scores + signals
├── report.md            # human-readable summary + bucket counts
├── health.json          # data-freshness + source counts
└── letters/             # populated by stage 2
    └── <address>.md     # one draft per top-N candidate

Listing-change monitoring

Re-scraping a listing to check for a price drop or delisting doesn't need a full re-read — diff the snapshot instead.

• Snapshot scraped listing content as markdown on disk, keyed by URL + date: snapshots/<url-hash>/<date>.md.
• On re-scrape, run Python's difflib.unified_diff against the most recent snapshot and surface only the changed hunks — a price drop, a status change (Active → Pending / Off Market), a delisting.
• Skip snapshots that produce zero diff; only act on candidates whose listing actually moved.

import difflib
old = open(f"snapshots/{url_hash}/{prev_date}.md").readlines()
new = open(f"snapshots/{url_hash}/{today}.md").readlines()
diff = list(difflib.unified_diff(old, new, lineterm=""))
if diff:
    print(f"{url}: changed\n" + "\n".join(diff))

No embeddings, no LLM call per re-scrape — a plain unified diff is enough to catch the signals that matter (price, status, delisting) without re-reading the whole listing each time.

(Pattern source: firecrawl's change-tracking feature — plain unified diff over markdown snapshots, apps/api/src/services/monitoring/diff.ts. This is a from-scratch local version, no API dependency.)

Counties supported

• allegheny_pa — Allegheny County, PA (Pittsburgh and inner suburbs). Source: WPRDC (Western PA Regional Data Center).

More counties via plug-in adapters — see references/adding-county-adapter.md for the checklist. A new county needs a single Python file in scripts/county_adapters/ exposing fetch_parcels(), fetch_sales(), fetch_tax_liens(), and fetch_sheriff_sales().