The Deal Database That Scans Greece While I Sleep

Direct Answer

carvana-ai is a Greek AI deal scout: every morning it scans Car.gr / AutoTriti (and optionally Facebook Marketplace view-only in your Chrome), saves listings to SQLite, values them with Greek comps (not KBB), and texts you Telegram cards with drafts you approve and send. Same Carvana economics (buy under market → sell higher): different tools, different trust model.

Key Takeaways

• The viral reel (Marketplace MCP + KBB + auto-DM) is a US play; it breaks in Greece.
• Start with one database and one search: not a mega-prompt.
• Pipeline: /car-scan → /car-value → /car-outreach → Telegram.
• Flag when ≥15% under Greek median, positive margin, ≥6 comps.
• The agent never messages sellers; Telegram drafts, your thumb on send.

Ampas · Thessaloniki · Episode 11 energy, Greek edition

---

Coders are looting the Carvana playbook. Look up what Carvana actually does: buy a car, pick it up, wash it down, maybe fix something, maybe sell it for on average seven grand more. Episode 11 of the AI-hustle internet says: we just do this with AI. Make a list of cars. Connect your agent to a Facebook Marketplace MCP. Paste one mega-prompt. Every morning it searches Marketplace, checks Kelley Blue Book, and literally messages sellers on your behalf. Texts you when it finds a banging deal. Turn the steps into skills. Schedule them. Don’t pay an extra $7,000 for a car.

Cute reel. Wrong country. Wrong valuation. Wrong trust model.

KBB prices America. Your buyer is in Ελλάδα. A Facebook Marketplace MCP that DMs strangers is how you get banned, not how you flip a Golf 7 in Θεσσαλονίκη. So we kept the Carvana economics: buy under market, light work, sell higher: and rebuilt the machine for the Greek used-car market: Car.gr + AutoTriti (+ Marketplace view-only in your Chrome), Greek comps instead of KBB, Telegram drafts you send by hand.

That’s carvana-ai. Same hustle. Not the same prompt. Copy-paste prompts and real code below.

---

The viral prompt vs what we actually run

PROMPT 01 · THE REEL (US fantasy: do not run this in Greece)

Paste into your agent. Swap the {slots}.

1. Every morning check Facebook Marketplace for these searches: {my car list}.
2. Save every new listing: price, mileage, year, trim, photos, location, seller: to my database.
3. For each new car, look up its Kelley Blue Book private party value and flag anything more than {15%} under.
4. For anything flagged, message the seller, ask if it is still available and why they are selling, and negotiate toward {my max}.
5. Text me only when you need a decision or the seller says yes.
6. Save this as three skills and run them on a schedule.

What’s wrong for us: KBB ≠ Greek market. Auto-DM ≠ allowed. Marketplace MCP scrapers ≠: we only view FB in your logged-in Chrome.

PROMPT 01b · THE WHOLE THING (Greek reality: paste this)

Paste into Claude / Cursor. Swap the {slots}.

Every morning run my Greek car-flip scout (carvana-ai):

1. Search Car.gr + AutoTriti (and Facebook Marketplace Ελλάδα ONLY via my
   logged-in Chrome, view-only) for: {my car list from config/searches.json}.
2. Save every new listing: price, mileage, year, fuel, transmission,
   engine_cc, trim, photos, location, seller_type: to data/listings.db.
3. Value each car with GREEK MARKET COMPS (same make/model/fuel/transmission,
   year ±2, km ±30.000, engine_cc ±200). Flag anything ≥ {15%} under the
   Greek median with positive margin after μεταβίβαση + transport + repair reserve.
   Do NOT use KBB.
4. For anything flagged, DRAFT a short Greek seller message (no price in the
   opener) and send ME a Telegram card. I approve, I copy, I send manually,
   I tap «Το έστειλα». Never message sellers yourself.
5. Text me only when I need a decision or a seller says yes.
6. Use skills /car-scan, /car-value, /car-outreach: orchestrate with /car-daily.

Then turn each step into a skill and schedule /car-daily.

---

1️⃣ Brain-dump what the AI can’t see

Open a note. Dump everything: models you’d flip, budgets, max offers, home base, which cities are “local,” fees, red flags, sellers you’ve already hit.

PROMPT 02 · ARCHITECT THE TABLES

You are my AI systems architect for a Greek used-car flipping scout.
Home base: {Θεσσαλονίκη}. Currency: EUR.
Sources: Car.gr, AutoTriti, optional Facebook Marketplace (Chrome view-only).

Brain-dump:
{paste your full list}

Hard rules:
- Never message sellers automatically
- Never buy, never handle deposits
- Valuation = Greek comps only (no KBB)

Do this:
1. Cluster into candidate database tables (name, columns, row shape).
2. Rank by leverage-to-effort.
3. Tell me the ONE table I should build first: defend it in two sentences.
4. Ask anything you need before that recommendation.

Do not write SQL yet.

Spoiler: first table is listings.

---

2️⃣ See the machine: ingestion · database · agent

/car-scan  →  /car-value  →  /car-outreach  →  Telegram
   │              │               │
   └──────── data/listings.db ────┘

Deal lifecycle:

new → alerted → approved|skipped → messaged → negotiating → accepted|dead

---

3️⃣ Start with the database (and ONE search)

Code · boot the schema

python scripts/init_db.py

Code · your car list (config/searches.json)

{
  "name": "VW Golf 7 2013-2019 βενζίνη",
  "keywords": "volkswagen golf",
  "make": "volkswagen",
  "model": "golf",
  "year_min": 2013,
  "year_max": 2019,
  "fuel": "βενζίνη",
  "max_mileage_km": 160000,
  "max_price_eur": 13000,
  "max_offer_eur": 10500
}

Code · secrets (.env)

TELEGRAM_BOT_TOKEN=123456789:AAAA-your-token-here
TELEGRAM_CHAT_ID=123456789

Create the bot with @BotFather → /newbot. Then:

python scripts/telegram.py chat-id

---

4️⃣ Wire the agent: commands you’ll actually use

python scripts/init_db.py
python scripts/db.py upsert path\to.json
python scripts/db.py unvalued
python scripts/comps.py run
python scripts/db.py deals new
python scripts/db.py summary
python scripts/telegram.py test
python scripts/telegram.py send-deal ID
python scripts/telegram.py poll
python scripts/telegram.py digest "Morning scan: …"

All scripts print JSON. Stdlib only. Python 3.14.

---

5️⃣ Prompts that make Claude the engineer

PROMPT 03 · DESIGN THE SCHEMA (don’t create yet)

I want tables in data/listings.db for a Greek car-flip scout:
listings, valuations, deals, messages.

Do NOT create them yet: design first as a senior database engineer:
- every column + type + why
- PKs, FKs, cardinality
- indexes for "unvalued listings" and "deals in state X"
- what you'd simplify for a single-user SQLite file vs production Postgres

Show proposed DDL, walk me through choices, wait for my go-ahead.

PROMPT 04 · CREATE + SAMPLE ROWS

Create the tables we designed using python scripts/init_db.py.
Confirm each step. Then insert two realistic sample Golf 7 listings (EUR, Greek
cities) and show me SELECT output so I can see it working.

PROMPT 05 · RESEARCH GREEK COMPS (not KBB)

Research why Kelley Blue Book is the wrong valuation source for flipping cars
in Greece. Then design a comps model that uses only listings we already stored:
match rules, median + p25, discount, local vs national tier, costs, flag when
≥15% under median with positive margin and ≥6 comps. Do not create code until I say go.

PROMPT 06 · BUILD THE MORNING SKILLS

Create (or refine) four Claude skills:
- /car-scan  : browser ingest Car.gr + AutoTriti (+ optional FB view-only)
- /car-value : run python scripts/comps.py run, sanity-check flagged pages
- /car-outreach: Telegram cards; I send every seller message manually
- /car-daily : orchestrate all three + greek digest

Hard rules stay. Walk me through install/use.

---

6️⃣ Automate data in: two lanes

Lane A: market (browser):

python scripts/db.py upsert data\scan_batch.json
python scripts/comps.py run

Lane B: your head (Telegram): approve drafts, paste seller replies, mark sent.

PROMPT 07 · LIVE TEST UNTIL A ROW EXISTS

Using /car-scan, pull the first page of results for my Golf search on Car.gr,
upsert into listings.db, run comps.py, and do NOT say you're done until
python scripts/db.py summary shows new rows and any flags. Fix parse issues
yourself. Then show me the JSON summary and one sample listing.

---

7️⃣ Make it DO things (skills > one viral prompt)

deal if:
  comps >= 6
  AND discount_pct >= 15
  AND est_margin > 0
  AND within search max_price / max_km / fuel

year ±2 · mileage ±30_000 km · engine_cc ±200
+ same make/model · same fuel/transmission when known

transfer_fee_eur: 200
repair_reserve_pct: 0.05
national_transport_eur: 300

PROMPT 08 · DEAL RANKING SKILL

I have listings, valuations, and deals in data/listings.db.
Create a skill /car-rank that pulls today's flagged deals, ranks by
(est_margin * confidence), separates local vs national, and returns a
Telegram-ready digest in Greek. Save it so I can re-run by name.

---

8️⃣ Guardrails (the part the reel skips)

• Agent never messages sellers. You copy from Telegram.
• No Facebook scraper MCPs, headless FB, or cookie export.
• No purchases, deposits, or payment details in drafts.
• Drafts never exceed max_offer.
• Pipeline stops at “seller said yes” → you do inspection, ΚΤΕΟ, transfer.

You own the data. The agent just visits.

---

Summary

The reel was half-right: context + skills + schedule beats prompt packs. It was half-wrong: KBB + auto-DM is US cosplay. In Greece you compound a listings table, value with Greek comps, and keep your thumb on send.

python scripts/init_db.py

Then tell your agent:

/car-daily

One question: if your AI could see ONE slice of the Greek used-car market tomorrow morning: which model is it?