Listing Scrape

Listing Scrape

Fetch a rental listing and turn it into facts.

What this skill is, and is not

This skill gets a listing page and extracts what it says. That is the whole job: fetching, parsing, and recording, including the parts that are hard because sites defend themselves and marketing copy lies.

It does not judge the listing. Gates, scores, statuses, decision rules, and where a row goes on a board all belong to the consuming repo's own scoring or playbook docs, and none of it is repeated here.

Where things go

This skill is shared across repos and does not know any repo's layout. Before writing anything:

1. Look for the repo's own conventions — a playbook or scoring doc, a requirements spec, a prospect template, an existing prospect file to pattern-match against. Follow what you find rather than improvising.
2. If the repo defines a prospect template, use it. This skill deliberately does not carry one.
3. If nothing in the repo says where prospect files live, ask rather than inventing a path.

Read those docs at the point you judge the listing, and follow them rather than any remembered version. A stale copy of a rule is worse than a pointer to it — gates and rent bases change often.

Fetch method

Try markdown first, fall back to HTML.

Step 1 — markdown:

lightpanda fetch '<url>' --dump markdown --wait-until networkidle

Step 2 — if step 1 returns no content, fall back to HTML:

lightpanda fetch '<url>' --dump html --wait-ms 5000 2>/dev/null

Read the raw HTML and pull the listing out of wherever it lives: JSON-LD (RealEstateListing or Product schema, usually the richest source on these sites), meta tags, __NEXT_DATA__ or similar hydration blobs, or body text.

When the fetch fails

Real estate sites defend themselves harder than job boards do, so expect this to fail more often than a careers page. Known behaviour:

• Realtor.ca (CREA) blocks automated clients aggressively and its terms prohibit scraping. Assume it will not work and do not build retries around it.
• HouseSigma and Condos.ca require login for most detail and sit behind bot protection.
• Rentals.ca, Zumper, PadMapper, and individual building sites are the ones most likely to actually return content.
• Facebook Marketplace needs a session and will not work.

Do not escalate. No proxies, no user-agent spoofing, no retry loops, no bulk crawling of search-result pages. This skill is for pulling one listing already being looked at, which is ordinary personal use. Grinding through a site's defences or harvesting listings in bulk is not, and it is how an account or an IP gets banned in the middle of a search.

When a fetch fails, say so plainly and ask for the listing text or a screenshot instead. Everything below the fetch works identically on pasted content, so this is a mild inconvenience rather than a dead end.

What to extract

Pull whatever the page actually has. Never infer a number that is not stated: unknown is a valid and useful answer, and a wrong square footage is worse than a blank one.

• Address, building name, neighbourhood, unit number
• Advertised rent, and what it includes (hydro, water, heat, internet)
• Interior square footage, excluding balcony and terrace. If the page gives one combined number, record it and flag it as unverified
• Bedrooms and bathrooms, plus the exact wording ("2 bed", "1+den", "2BR + flex")
• Parking: included, available at a cost, or none. Note motorcycle handling if mentioned
• Locker or storage
• Year built
• Building amenities, with the gym called out specifically
• Laundry, balcony, outdoor space, orientation and exposure
• Availability date, lease term, pet policy
• Listing agent or landlord, and whether it is an individual or a management company

Two rules that are about scraping, not judgment

These stay here because they govern what counts as a fact, which is this skill's problem:

• Marketing copy is not evidence. "Fitness centre", "spacious den", "800 +/-", "steps to everything" record as what they are — claims — and never as confirmed facts. Only a stated, specific figure or an explicit inclusion counts.
• Record the exact wording for anything ambiguous: bedrooms, parking, square footage, what the rent includes. The judgment gets made later against what the listing actually said, not against a paraphrase.

Output

Write a prospect file using the repo's template, filling the factual sections and turning the gaps into specific unanswered questions so they actually get asked at a viewing. Leave blank whatever the page did not say. A blank is a useful signal; an invented number is not.

Then hand off: read the repo's scoring or playbook doc and follow it to fill the gate line, the scores, the status, and any roster row. That step is governed there, not here.

Report back

Short. Do not restate the file.

• Address, advertised rent and what it includes, square footage, bedrooms as worded, parking, locker, year built
• What the page did not say, which is usually fibre, the gym equipment, and a real interior square footage
• Anything the listing contradicted itself on. The site's data tables and its descriptions disagree often enough to be worth flagging every time
• Then the gate line, score, and verdict as the repo's scoring doc defines them

Hold whatever reporting guardrails the repo's CLAUDE.md sets when reporting.