flip-deal-feed

Sweep live residential and small-commercial listings in a target buy-box, screen every one through the 70% flip rule, firm ARV from real sold comps, and surface a ranked shortlist with first-pass deal

SKILL.md

GitHub

Flip Deal Feed

Turn a whole metro's listings into a handful of properties worth driving to. Given a buy-box (cities + drive-time constraint) and a budget, this sweeps the live listings, screens each through the 70% rule, estimates ARV from real sold comps, and ranks by margin of safety — then writes a practical, no-nonsense brief for the survivors.

Reference run (2026): a five-suburb inner-ring buy-box, 276 listings swept, 5 prospects surfaced. The screen/calculator engine is scripts/flip_screen.py; the commercial underwrite is scripts/cre_underwrite.py. Both run standalone (python3 flip_screen.py prints a demo on fictional listings).

Write plainly (see copy-voice). The brief is for someone evaluating a buy — practical and specific (what needs doing, what to offer, what to confirm onsite), never hype.

The honest tooling split (read first)

  • Sweeping listings = WebFetch the listing pages. NOT Gemini. The pages below render server-side, so the native WebFetch tool retrieves them cleanly. (Redfin, Zillow, homes.com, Movoto, LoopNet, Crexi all 403 automated fetches — don't bother.)
  • ARV + comps + rents = grounded web search (gemini-web-search). Finding recent sold comps near an address, or a commercial building's submarket rents, is a research question. Using search to fetch a known listing URL is the wrong tool.

Step 1 — Sweep the buy-box (WebFetch)

Residential — Century21 (renders, returns structured rows):

https://www.century21.com/real-estate/<city>-<st>/LC<ST><CITYUPPERNOSPACES>/
# e.g. springfield-il/LCILSPRINGFIELD/  (state code lower in the slug, upper in the id)

Fetch one city per call; ask for address, price, beds, baths, sqft; exclude condos/townhomes. A city page returns ~50–75 single-family rows. Sweep every city in the box.

Commercial — CommercialSearch (renders; the residential sibling point2homes 403s):

https://www.commercialsearch.com/for-sale/commercial-real-estate/us/<st>/<city>/

Ask for address, price, type, building sqft, units, and any cap/NOI. Skip vacant land. But CommercialSearch is thin — most box inventory is over-budget, "contact for pricing," or missing NOI. For the small multifamily value-add plays (4–16 units, the real commercial fit for "renovate → raise rents → resell"), use grounded search — it reads Google's index and surfaces Crexi/Zillow/LoopNet listings (with address, price, units, even the in-place NOI) that all 403 a direct fetch:

python3 gemini_search.py "small apartment buildings 4-16 units for sale under \$1.5M in <cities>, with address, price, units, NOI"

Commercial valuation (NOI ÷ cap), flip_screen.py:

from flip_screen import value_add
value_add(noi_now, noi_stabilized, market_cap=0.0575, price=795_000)
# going-in cap, in-place value at market cap, stabilized value, value created

Value = NOI ÷ cap, not a comp spread. The value-add lever is below-market rents + trimmed opex; every $1 of NOI growth = $1 ÷ cap of value. Commercial is a 12–24mo reposition, not a 6-month flip — say so; the timeline differs from residential.

Full levered underwrite (DSCR, cash-on-cash, IRR, exit at a cap range): scripts/cre_underwrite.py — underwrite(price, noi_in, noi_stab, capex, down=, rate=, exit_cap=). Models a value-add ramp (negative year-1 carry for a vacant-on-purchase deal). Always run the buy-below-ask + cheaper-debt scenario — at ask it usually looks like a dud.

Step 2 — Screen through the 70% rule (flip_screen.py)

from flip_screen import screen
rows = screen(listings, arv_by_city)   # ranked; clears_at_ask flag + discount_to_offer_pct
  • MAO = 0.70 × ARV − rehab. The 30% gap is financing + holding + selling + profit.
  • Rehab is a uniform desk estimate ($/sqft); the brief refines it per property.
  • The deals are the listings priced below the area cluster. At-retail listings won't clear — that's normal. Surface the mispriced/distressed ones and the value-add plays (2/1 → 3/2, add a bath, finish a basement).

Step 3 — Firm the ARV (grounded search)

For each survivor, pull real recent sold comps to replace the desk ARV:

python3 gemini_search.py "recent sold comparable homes near <address>, renovated, <beds>bd <sqft>sqft, last 6 months, with sale prices"

Show ARV as a range, not a point (flip_screen.deal() returns net upside across an ±8% ARV band). A flip lives or dies on ARV − all-in; never print a fake-precise profit.

Step 4 — The deal math (flip_screen.deal())

Generic, no lender required — vary down% and rate to see the levers:

from flip_screen import deal
deal(purchase=160_000, rehab=50_000, arv=320_000, down=0.20, rate=0.08, months=6)
# -> cash_in, monthly_payment, net_low/mid/high

Report the three numbers the buyer asks for: cash down needed · monthly payment · net upside (range).

Step 5 — The prospect brief (the deliverable)

One page per survivor, built from data we have, with the unknowables flagged:

  1. Snapshot — address, ask, beds/baths/sqft, year built, last sale, assessed value, days-on-market (county record + listing; all real).
  2. Buy-right number — MAO vs ask. "Offer $X. Asking $Y. Gap $Z." The go/no-go.
  3. ARV — 3–5 real sold comps (grounded search), as a range.
  4. Rehab scope — line items from photos/age with $ ranges, and explicit CONFIRM ONSITE for what a desk can't see (foundation, sewer line — get a $150 scope, panel, asbestos/lead pre-1978, roof + mechanicals age). Note where the buyer's own sweat labor halves a cosmetic line, but keep licensed trades at full price.
  5. Deal math — cash in, monthly, net upside range.
  6. Verdict — GO / NO-GO / WALK IT.

Buy-box / drive-time

Confirm each candidate against the drive-time constraint before it counts (e.g. within 20 minutes of two anchor addresses). A cheap house outside the box is not a prospect.

Learnings / guardrails (earned on the reference run — don't relearn them)

  1. A cheap attached unit is not a discounted house. Century21 sweeps include townhomes/condos. A "$160K, 3/2" looked like a 50%-off distress vs a $320K SFR ARV — it was a 1983 townhome whose portal AVM ($160,431) = the list price. No discount; priced right. Verify property TYPE and cross-check list price against an AVM (Redfin/Zillow estimate) before calling anything a discount. flip_screen.screen() now flags type in {townhome,condo} and ask >= 0.93*avm.
  2. Verify ACTIVE status before briefing. A 4-plex had the biggest rent gap on the board — and was Pending. Search results can be stale (it surfaced a 2024 status). Confirm the listing is live, not pending/sold.
  3. Opex structure is the value driver, not just rents. An 8-unit (owner pays heat+water+trash → 46% opex) vs a 7-unit (electric separately metered, owner pays only water/sewer → ~35%). Capture who pays which utility — master- metered heat/water is the RUBS/sub-meter upside, often bigger than the rent bump.
  4. Use the right cap rate for the class/era. 1950s–60s Class-C walk-ups traded ~6.5–8% in the reference market, NOT the 5.6–5.75% institutional Class-B figure. Using the low cap overstated value-creation ~2× on the first pass. Show a cap-rate football field (value at 6.0/6.5/7.0/7.5%) — never a single cap.
  5. Run the going-in DSCR. At ~7% commercial debt, these buildings' in-place NOI barely covers debt service (DSCR ~1.03 at 25% down). The deal only becomes financeable after the value-add — plan 30%+ down or a buy-down, and say so.
  6. Strategic read for that box (check yours): SFR flips are priced efficiently (retail) — the survivors are mostly cheap attached units with no real margin. The edge here is small-multifamily value-add, not house flips. Steer effort accordingly.
  7. Negative leverage is the trap whenever cap < debt rate (it was in 2026). These small-MF/industrial deals list at 6–6.5% caps while debt is ~7% — cap < rate, so leverage drags returns down. At ask, all three live prospects underwrote to 1–4% levered IRR. They only reach a real return (12%+) by buying ~12% under ask (push going-in cap above debt cost) **+ cheaper agency/credit-union debt (6.5%).** Always lead the underwrite with that scenario; never present an at-ask IRR as the deal. The money is made on price, not reno.

Done when

  • The whole box is swept (every city), not one city — state the count evaluated.
  • Survivors are ranked by margin of safety, ARVs are real comps (not desk estimates), and each has a target offer.
  • Every figure is real public data or a clearly-labeled estimate — nothing invented.
  • The brief reads practical and specific (what to offer, what to fix, what to confirm onsite), plain language holds.
  • Commercial survivors carry an NOI/rent-roll caveat if the listing didn't publish one (value = NOI ÷ cap).