yelp（yelp 搜索）

When to Use

User wants to work on Yelp directly: find restaurants or local services, compare businesses, inspect reviews, check delivery or takeout signals, or audit a Yelp listing with source-aware workflows.

Use this skill when the answer depends on Yelp-specific fields such as rating, review count, categories, price, location, transactions, hours, photos, or business-page evidence rather than generic search alone.

Architecture

Memory lives in ~/yelp/. If ~/yelp/ does not exist, run setup.md. See memory-template.md for structure.

~/yelp/
├── memory.md                 # Activation behavior, preferred markets, and decision patterns
├── sessions/
│   └── YYYY-MM-DD.md         # Current task context and selected candidates
├── businesses/
│   └── {city-or-segment}.md  # Saved shortlists, business aliases, and notes
├── api/
│   ├── alias-cache.md        # name+location -> Yelp alias or business ID
│   └── request-log.md        # Redacted endpoint, params, status, timestamp
└── audits/
    └── {business}.md         # Listing-quality findings and next actions

Quick Reference

Load only the file needed for the current Yelp task.

Requirements

• No credentials required for public-page verification and manual comparison.
• YELP_API_KEY is recommended for official API search, details, and reviews workflows.
• Owner-side or partner-managed tasks require explicit user approval and the user's own authorized access.
• Confirm before sending phone numbers, exact addresses, or account-scoped listing data to live Yelp endpoints.

Coverage

This skill is designed for Yelp-specific work that usually breaks when an agent treats it like generic local search:

• restaurant and local-service discovery
• head-to-head business comparison in one market
• review signal analysis and complaint clustering
• delivery or takeout eligibility checks where Yelp exposes those signals
• business listing audits for owners, operators, or agencies

Core Rules

1. Choose the operating mode before querying

• Use page mode for visible public research, API mode for structured fields, and audit mode for listing-quality reviews.
• State the chosen mode in the output so the user knows where the evidence came from.
• If API access is missing or degraded, continue in page mode and explain the tradeoff.

2. Resolve the exact business before analysis

• Use location, category, alias, and phone match when possible before comparing or quoting review signals.
• Chains and duplicate names often produce wrong merges across neighborhoods or cities.
• If identity remains ambiguous, ask one disambiguation question instead of guessing.

3. Separate consumer research from owner actions

• Default to read-first workflows: search, compare, summarize, and audit.
• Draft review replies, escalation notes, or listing changes only if the user is authorized and asks for them.
• Do not imply account access, verified ownership, or live edit capability without explicit proof.

4. Rank with decision signals, not stars alone

• Weight rating, review volume, review recency, complaint themes, price fit, and category match together.
• Recent negative clusters often matter more than a legacy 4.5 average.
• Flag low-sample or stale-review situations clearly.

5. Treat Yelp fields as platform-specific, not universal truth

• price, transactions, hours, attributes, and delivery or takeout flags are useful but not guaranteed to be current everywhere.
• Re-check operational claims close to the decision when the user is about to book, visit, call, or order.
• Do not convert Yelp metadata into hard promises.

6. Keep storage minimal and redacted

• Store reusable filters, accepted or rejected shortlist reasons, and verified aliases in ~/yelp/.
• Never store secrets, raw API keys, or unredacted signed request URLs.
• In logs, keep only endpoint path, safe params, outcome, and timestamp.

7. Return outcome-ready outputs

• For discovery: give a shortlist with why each option is in, what the tradeoffs are, and what should be verified next.
• For audits: group issues into highest-impact fixes first, then supporting cleanup.
• For review analysis: separate evidence, inference, and uncertainty.

Common Traps

• Comparing businesses from different neighborhoods without distance or market normalization -> the shortlist looks better on paper than in real life.
• Treating a high rating with very low review volume as stable -> one or two new reviews can flip the signal.
• Ignoring the newest negative reviews -> current service decline stays hidden behind historical averages.
• Assuming delivery, takeout, or hours flags are universally current -> users act on stale operational data.
• Mixing consumer lookup with owner-side tasks -> the agent drifts into unauthorized or impossible actions.
• Logging raw API request headers or URLs -> secrets leak into local notes.

External Endpoints

No other data is sent externally unless the user explicitly approves another source.

Security & Privacy

Data that may leave your machine:

• business names, search terms, location hints, phone numbers, and optional filter parameters sent to Yelp

Data that stays local:

• shortlist decisions, recurring filters, alias cache, and audit notes in ~/yelp/

This skill does NOT:

• store API keys in markdown files
• claim listing ownership or edit access without proof
• use CAPTCHA bypass, anti-bot evasion, or undeclared scraping flows
• modify its own SKILL.md

Trust

This skill can send business search terms, locations, and optional identifiers to Yelp services. Only use live Yelp calls if you trust Yelp with the local-search context relevant to the task.

Scope

This skill ONLY:

• searches and compares Yelp businesses with explicit evidence
• analyzes review signals, listing quality, and operational flags
• prepares drafts or audit recommendations when the user is authorized

This skill NEVER:

• fabricate ratings, reviews, or business availability
• imply owner access or post live changes without explicit instruction
• hide missing data, stale signals, or unsupported market coverage

Related Skills

Install with clawhub install <slug> if user confirms:

• google-reviews - Compare Yelp signals against Google review patterns and reputation drift.
• restaurants - Turn shortlisted places into actual dining recommendations and decision filters.
• maps - Add routing, geocoding, and distance checks before the user acts on a shortlist.
• apple-maps - Open chosen places quickly on macOS for directions and location confirmation.
• tripadvisor - Cross-check travel-heavy restaurant and attraction choices outside Yelp.

Feedback

• If useful: clawhub star yelp
• Stay updated: clawhub sync