Build a research assistant from sample listings
https://patrickz.ai/learn/listing-research-agent/

OUTCOME
Normalize and compare listings without silently inventing missing facts.

YOU NEED
A chat or coding assistant and five saved fictional rental listings. No scraping or messaging is required.

1. Define the practice version
Define a common schema, a stable source identifier and explicit unknown fields. Start with saved examples before connecting a live website.

2. Build step 1
Define a listing schema before scraping.

3. Build step 2
Start with saved sample data.

4. Build step 3
Build scoring and filters against those samples.

5. Build step 4
Add SQLite persistence and deduplication.

6. Build step 5
Implement one site adapter.

7. Build step 6
Add a desktop table and detail panel.

8. Build step 7
Generate outreach drafts, but require a human click to send.

9. Run the experiment
Include one duplicate, one stale listing and one conflicting price. Confirm the output preserves the original values and explains why each item was flagged.

PROMPT
Work in a disposable practice project. Explain any setup requirements before changing files. Build one small step at a time and show how I can check it.

Design a normalized rental-listing schema for multiple websites. Include source ID, canonical URL, project, price, size, beds, baths, location, transit, amenities, agent contact, listing date, extraction confidence, and conflict flags.

Create an explainable scoring rubric for Bangkok apartments. Rank budget, accessibility, minimum size, transit, parks, furnishing, stale dates, and contradictory fields. Every score must produce human-readable reasons.

My first-version boundary: Define a common schema, a stable source identifier and explicit unknown fields. Start with saved examples before connecting a live website.

EXAMPLE / EXPERIMENT
Include one duplicate, one stale listing and one conflicting price. Confirm the output preserves the original values and explains why each item was flagged.

CHECK YOUR RESULT
[ ] Duplicates are identified without losing source links.
[ ] Missing and conflicting values remain visible.
[ ] Any outreach remains a human-reviewed draft.

IF IT FAILS
Do not assume a publicly viewable page permits automated collection. When adding a real source, check its access rules and use an approved API or permitted workflow.

MY RESULT / NEXT CHANGE