City Rental Hunt#

Turn fuzzy apartment-hunting requests into a repeatable workflow:

1. normalize constraints
2. generate zone-aware search keywords
3. search social listings
4. extract listing facts
5. filter red flags
6. produce a shortlist and contact brief

Quick start#

When the user asks to hunt rentals in a Chinese city:

1. Normalize requirements first
   • city
   • target zones
   • budget range
   • room count
   • must-have constraints
   • one-vote vetoes
2. Generate search keywords

   • Run: bash

     python3 skills/city-rental-hunt/scripts/keyword_plan.py \
       --city 北京 \
       --zones "北苑,霍营,清河" \
       --budget "6000-9000" \
       --rooms "两居" \
       --must "整租,电梯,次新" \
       --optional "可养猫,房东直租,转租"
     
     

   • This produces reusable search phrases for each zone.

3. Search platforms in this order
   • First: Xiaohongshu via TikHub
   • Second: Douyin via TikHub
   • Use the existing social-media skill/tooling instead of inventing new scraping flows.
4. Collect only listing-relevant facts
   • platform
   • post id / URL
   • title / short summary
   • price if present
   • neighborhood / subway / zone
   • freshness
   • landlord / agent / unclear
   • pet policy if present
   • likely keep / maybe / discard
5. Output a shortlist, not a dump
   • Keep the result decision-oriented.
   • Separate high-confidence leads from weak leads.

Workflow#

Step 1 — Normalize the requirement brief#

Use this compact schema:

yaml

city: 北京
zones: [北苑, 霍营, 清河]
budget: 6000-9000
rooms: 两居
must_have:
  - 整租
  - 电梯
  - 次新/不要老小区
soft_preferences:
  - 客厅大
  - 房东直租
  - 靠近地铁
  - 宠物友好
vetoes:
  - 老小区
  - 合租
  - 商住
  - 非民水民电

If the user gives vague input, infer only the search structure, not the final preference.

Step 2 — Build search buckets by zone#

Do not search one giant keyword first. Split by zone.

For each zone, create 3 buckets:

1. broad: 北苑 整租 两居
2. quality: 北苑 次新 电梯 两居
3. conversion: 北苑 房东直租 两居 / 北苑 转租 两居 / 北苑 可养猫

If a known neighborhood appears repeatedly, promote it into its own bucket.

Step 3 — Search Xiaohongshu first#

Use TikHub endpoints exposed by the existing social-media skill. Typical flow:

• check help
• check list-endpoints xiaohongshu when needed
• search notes with short, high-signal phrases

Prefer short Chinese queries over long natural-language queries. TikHub/XHS search often degrades on long keyword strings.

Step 4 — Search Douyin as a supplement#

Use Douyin only after XHS has produced a first-pass pool.

Douyin is useful for:

• video walk-throughs
• “刚空出来” style posts
• transfer/转租 leads

Do not let Douyin dominate the run unless XHS is thin in that city/zone.

Step 5 — Extract and classify leads#

For every lead, classify:

• keep: fresh, plausibly matches constraints, enough information to contact
• maybe: missing price / pet policy / building age, but still promising
• discard: clear red flag

Use the red-flag checklist in references/playbook.md.

Step 6 — Produce two outputs#

Output A: analyst-facing search record

Include:

• keywords used
• leads found
• keep/maybe/discard reasoning
• repeated neighborhoods worth deeper follow-up

Output B: user-facing morning brief

Include only:

• top leads
• why they matter
• what to contact first
• key uncertainties to verify

Scoring heuristics#

Use these dimensions:

• freshness: today / yesterday / within 7 days / stale
• constraint fit: rooms, budget, elevator, new-enough community
• contactability: landlord direct > personal transfer > unclear > obvious agent spam
• risk: old community, no elevator, price missing, ad tone, commercial apartment, shared rental smell
• special upside: pet-friendly, unusually concrete price, exact move-in date, strong transit fit

A listing with incomplete price can still rank high if it is very fresh and structurally fits.

What to avoid#

• Do not treat every social post as a real listing.
• Do not present stale posts as active inventory.
• Do not bury the user in 30 weak links.
• Do not confuse “cheap” with “good fit”.
• Do not publish or embed private commute addresses or personal names when turning a private search workflow into a reusable skill.

Default report shape#

Use this structure unless the user asks otherwise:

markdown

# Rental hunt brief

## Requirement snapshot

## Zones searched

## Top leads
- lead
- lead
- lead

## Backup leads

## Repeated neighborhoods worth deeper checking

## Risks / unknowns to verify
- price
- pet policy
- landlord vs agent
- building age / elevator

## Contact-first order
1. ...
2. ...
3. ...

When to read the reference#

Read references/playbook.md when you need:

• a fuller keyword-building pattern
• a reusable evidence schema
• a red-flag checklist
• a morning-brief template