yinren112/xhs-rental-scout

XHS Rental Scout

Turn public Xiaohongshu rental posts into a small, decision-ready shortlist. Optimize for evidence quality and avoided contact cost, not result count.

Keep the boundary

• Read public information only. Never message, comment, like, follow, call, add contacts, book, or pay.
• Never bypass login, CAPTCHA, verification, rate limits, or platform restrictions.
• Stop new online requests on authentication or verification failure. Keep completed evidence and report the blocker.
• Keep cookies, tokens, raw caches, personal requirements, and account lists outside the installed skill directory and outside public repositories.
• Treat public posts as leads, not verified inventory or market-wide data.

Prepare the request

Obtain or infer these essentials before searching:

• city or target area;
• monthly budget, including whether it is per unit or combined;
• commute destination and acceptable route;
• acceptable home types;
• hard rejects and must-haves.

Ask no more than two short questions at once when an essential cannot be inferred. Do not require a configuration file for a one-off search.

For repeated searches, create a user-owned workspace from assets/rental-profile.example.md; follow references/workspace-and-output.md. Never write runtime data into this skill folder.

Select a read backend

Read references/backend-routing.md before the first online request. Reuse a working, authenticated capability already available in the environment. Do not install or build a scraper when an existing connector, browser session, MCP, or CLI can read the needed public fields.

Run the evidence funnel

1. Load the user's profile and prior seen-notes.txt / account-risks.md when present.
2. Generate 6-12 focused queries from area, home type, personal-intent, move-in, and property terms. Avoid one giant generic query.
3. Search the first page of each query. Continue to another page only while it produces new plausible leads.
4. Deduplicate by note ID first, then flag probable reposts using author, property, price, title, and repeated images. Do not merge uncertain matches silently.
5. Reject only hard mismatches at card level. Keep missing-title cards for detail review when other signals are relevant.
6. Read listing details for plausible new leads. Separate confirmed facts, reasonable inferences, and unknowns.
7. Inspect author history only for remaining leads. Judge the posting pattern, not a single nickname or slogan.
8. Inspect comments and images only for finalists or when they can resolve a decision-critical gap.
9. Apply references/evidence-rules.md. Keep evidence grade, proximity/pairing grade, and recommended action as separate fields.
10. Produce the report with assets/report-template.md; persist state only after recording why each note was shortlisted, downgraded, or rejected.

Preserve source meaning

• Keep title, body, image, profile, and comment evidence distinguishable.
• Record price as listed monthly rent, suspected rent, non-rent charge, or missing. Never convert missing price into a cheap listing.
• Prefer visible image evidence when it conflicts with prose, but record the conflict.
• Treat “房东直租”, “无中介费”, and similar claims as clues, never proof.
• Say “public-post sample” or “listed price”; never claim citywide market price or completed transaction price from these posts.

Finish only with a usable handoff

Report:

• scope, date, queries, and source limitations;
• ranked candidates with links and evidence gaps;
• rejected or downgraded leads with reasons;
• blocked checks and login/verification state;
• the smallest next question worth asking for each finalist;
• which local state files were updated.

If no lead meets the bar, say so plainly. Do not loosen the standard to fill the report.