Dianping Restaurant Guide (大众点评探店指南)#

Make smarter dining decisions by cutting through Dianping's noise. Detect fake reviews, extract genuine sentiment signals, and get personalized restaurant recommendations with dish-level ordering guides.

Core Capabilities#

• Review authenticity analysis: Detect suspicious review patterns — timing clusters, user activity profiles, language similarity, and photo metadata
• Sentiment distillation: Filter out probable fake reviews, then aggregate genuine opinions across taste/environment/service/value
• Personalized ranking: Score restaurants against your preferences (cuisine, budget, occasion, distance)
• Dish-level recommendations: Not just which restaurant — which specific dishes to order, and which to avoid
• Exploration route planner: Optimize a multi-restaurant exploration itinerary by geography
• Polarization decoder: When reviews are split 50/50, analyze what each camp values and help you decide

Workflow (8 Steps)#

Step 1: Define Dining Context#

Input: User provides:

• Location: City + district/landmark (e.g., "上海静安寺", "北京三里屯")
• Cuisine preference: Type (川菜/日料/西餐/火锅...) or "any"
• Budget: Per-person range (e.g., ¥50-100, ¥200+)
• Occasion: Date / family / solo / group dinner / business
• Distance radius: Walking (1km) / short taxi (5km) / willing to travel

Output: Structured dining profile. Logic: Fill missing fields with sensible defaults (budget = ¥50-150, radius = 3km). Ask clarifying questions for ambiguous inputs.

Step 2: Candidate Discovery#

Input: Dining profile. Action: Search Dianping for restaurants matching criteria. Filters:

• Rating 3.5–5.0 (include mid-range; some great finds are under-rated)
• Review count ≥50 (sufficient sample)
• Within radius
• Cuisine match

Output: Candidate list (typically 10–30 restaurants). Logic: Deduplicate chain locations unless user prefers them. Track "hidden gems" (high real-review score despite moderate overall rating).

Step 3: Review Harvesting#

Input: Candidate restaurant list. Action: Scrape recent reviews (sample ≥30 per restaurant, prioritized by recency). Data extracted per review:

• Rating (overall + sub-scores)
• Text content + length
• Photos (count + metadata)
• Reviewer profile: registration date, review count, follower count, review frequency
• Timestamp

Output: Raw review corpus per restaurant. Logic: Respect Dianping rate limits. Cache for 1 hour.

Step 4: Fake Review Detection (The Differentiator)#

Input: Review corpus per restaurant. Detection signals (multi-dimensional scoring):

SignalRed Flag PatternWeightTime cluster≥5 reviews within 30 minHighReviewer ageAccount <7 days oldHighReview frequency>10 reviews/day across restaurantsMediumText similarityHigh cosine similarity with other reviewsHighPhoto EXIFIdentical camera/settings across "different" usersMediumGeneric language"环境不错""味道很好" without specificsLowRating patternAll 5-star, no detailLowReviewer historyOnly 5-star reviews, never criticalMedium

Output: Each review tagged: genuine | suspicious | likely_fake. Restaurant scored with authenticity index (0–100).

Logic: Ensemble scoring with weighted signals. Flag entire restaurant as "suspicious" if >40% reviews are likely_fake. This is the key differentiator from any existing tool.

Step 5: Sentiment Extraction (Genuine Reviews Only)#

Input: Filtered genuine reviews. Action: LLM-based sentiment extraction across 4 dimensions:

• Taste (口味): Dish-specific praise/criticism, freshness, authenticity
• Environment (环境): Noise level, seating comfort, decor, cleanliness
• Service (服务): Wait time, staff attitude, ordering convenience
• Value (性价比): Portion size vs price, "worth it" sentiment

Output: Structured sentiment summary per restaurant.

text

=== 老王川菜 (Rating: 4.3, Authenticity: 92%) ===
👍 Taste: Mapo Tofu (must-order, 38 mentions), 回锅肉 (authentic, 25 mentions)
👎 Taste: 水煮鱼 (bland, 7 complaints)
👍 Environment: Cozy but cramped (expect wait at peak)
👎 Service: Slow during dinner rush (12 mentions)
💰 Value: ¥65/person, "great value" (22 mentions vs 3 "overpriced")

Step 6: Personalized Ranking#

Input: Sentiment summaries + user dining profile. Scoring formula:

text

Score = w1 × cuisine_match + w2 × sentiment_score + w3 × value_score + w4 × occasion_fit + w5 × authenticity_bonus

Output: Ranked top 5–10 restaurants with:

1. Overall score and breakdown
2. Must-order dishes (top 3)
3. Avoid dishes (top 3)
4. Best time to go (based on wait-time mentions)
5. Occasion-fit note (e.g., "Great for date — quiet corner tables")

Step 7: Ordering Guide Generation#

Input: Top-ranked restaurant + dish sentiment data. Output: Ready-to-use ordering card:

text

=== 老王川菜 · 2人点菜指南 ===
预算: ¥120-150

必点 (Must Order):
☑ 麻婆豆腐 ¥28 — 38人推荐，招牌菜
☑ 回锅肉 ¥42 — 25人推荐，正宗

可点 (Recommended):
○ 蒜泥白肉 ¥32 — 18人推荐

避雷 (Skip):
✕ 水煮鱼 ¥88 — 7人评价"不入味"，性价比低

💡 人均参考: ¥60-75
🕐 建议时间: 11:30前或13:30后避开高峰

Logic: Adjust portion recommendations based on party size. For group dinners, suggest "大份" or double-ordering popular dishes.

Step 8: Optional — Exploration Route#

Input: Multiple restaurants worth visiting + starting location. Output: Geographically optimized route:

text

🍽️ 静安寺美食探店路线 (步行可达):
1. 老王川菜 (午餐, 11:30) → 步行3分钟
2. 弄堂小馄饨 (下午茶, 14:00) → 步行5分钟
3. 福1088 (晚餐, 18:00)

Logic: Group restaurants by proximity clusters. Factor in operating hours (some close between lunch and dinner).

Sample Prompts#

Prompt 1: Quick Recommendation#

User: "上海静安寺附近，想吃川菜，人均100左右，约会，推荐3家靠谱的" Expected Output: 3 ranked restaurants with authenticity scores, must-order dishes, date-suitability notes.

Prompt 2: Suspicious Restaurant Check#

User: "这家网红店评分4.8但我觉得不好吃，帮我看看评价是不是刷的 —— [Dianping URL]" Expected Output: Fake review detection report: what % of reviews look fake, specific suspicious patterns found, genuine review sentiment summary.

Prompt 3: Dish-Level Deep Dive#

User: "今晚去吃XX火锅，帮我看看哪些菜必点，哪些是雷" Expected Output: Detailed ordering guide with specific dish recommendations sourced from genuine reviews.

Prompt 4: Polarized Restaurant Decoder#

User: "XX餐厅评分3.8，有人狂推有人狂骂，到底是什么情况？" Expected Output: Analysis of both camps: "5-star reviewers love the authentic taste and value; 1-star reviewers complain about slow service and cramped space. Verdict: go for the food, not the ambiance."

Prompt 5: Multi-City Explorer#

User: "下周去成都出差3天，住春熙路附近，帮我规划一个探店路线，要正宗川菜和小吃" Expected Output: 3-day exploration route: lunch + dinner + snack recommendations, geographically clustered, with ordering guides.

Prompt 6: Group Dinner Planning#

User: "8个人聚餐，北京国贸附近，预算人均150-200，要有包间，适合聊天的" Expected Output: 5 restaurant options with private room availability notes, group-friendly dish combos, per-person budget breakdown.

Real Task Examples#

Example 1: Date Night Decision#

Scenario: User planning a third date, wants to impress without being too formal. Input: "上海法租界附近，人均200-300，约会，氛围好但不要太正式，最好是西餐或日料" Steps:

1. Profile: 法租界, ¥200-300, date, Western/Japanese
2. Discovery → 18 candidates, filtered to 12 after radius check
3. Review harvesting → 360+ reviews across candidates
4. Fake detection → 2 restaurants flagged (>35% suspicious), removed
5. Sentiment: focus on "romantic atmosphere" and "not too noisy"
6. Ranking: Top 5 with "date vibe" score
7. Ordering guides for top 3 Output: "Mr & Mrs Bund (authenticity 95%, romantic terrace), 鮨一 (94%, intimate counter seating), Osteria (91%, cozy Italian)"

Example 2: Tourist in New City#

Scenario: First time in Chengdu, wants authentic local food, not tourist traps. Input: "刚到成都，住宽窄巷子附近，想吃地道川菜和小吃，避开网红店，人均50-100" Steps:

1. Profile: 宽窄巷子, ¥50-100, solo/casual, authentic
2. Discovery → 25 candidates; immediately filter out 4.8+ perfect scores (tourism red flag)
3. Review harvesting → focus on reviews from accounts with local dining history
4. Fake detection → remove 6 restaurants with review clusters on weekends
5. Sentiment: prioritize "本地人""回头客""老店" mentions
6. "Hidden gem" bonus: moderate rating (3.8-4.3) but strong genuine sentiment
7. Route: 3-restaurant walking route from 宽窄巷子 Output: "陈麻婆豆腐 (老字号, 96% authentic), 小谭豆花 (local favorite, ¥30/person), 洞子口张老二凉粉 (street food gem)"

Example 3: Business Dinner#

Scenario: Hosting 6 clients, needs upscale private dining with guaranteed quality. Input: "北京朝阳区，商务宴请6人，人均400-600，要有包间，粤菜或淮扬菜，服务要专业" Steps:

1. Profile: 朝阳区, ¥400-600, business, private room, Cantonese/Huaiyang
2. Discovery → filter for "包厢""商务宴请" in tags
3. Review focus: service quality mentions, private room experience
4. Fake detection → high-end restaurants have fewer fake reviews but check for "探店博主" (paid influencer) patterns
5. Sentiment: service quality is weighted 2× for business dining
6. Dish recommendations: banquet-style combos, wine pairing suggestions
7. Booking tips: how far in advance to reserve Output: "利苑酒家 (service excellence, 98% authentic), 大董烤鸭 (impressive presentation, private rooms), 淮扬府 (elegant, excellent service)"

🚀 First-Success Path (3 Steps)#

1. Step 1: Say "推荐上海静安寺附近3家靠谱的川菜馆，人均100左右"
2. Step 2: Review the ranked list with authenticity scores and must-order dishes
3. Step 3: Pick a restaurant, get the ready-to-use ordering guide with specific dishes

Boundary Conditions#

ConditionBehavior<50 total reviews for a restaurantFlag as "insufficient data"; still show but with low-confidence warningAll reviews flagged as suspiciousMark restaurant "⚠️ High fake review risk"; still show genuine sentiment if anyUser location outside Dianping coverageInform; Dianping covers mainland China primarilyCuisine type has <3 candidatesExpand radius or suggest adjacent cuisinesBudget too low for area (<¥30/person)Suggest street food / food court alternativesRestaurant chain with >10 locationsDe-duplicate; show 2-3 closest locationsDianping blocks scrapingFall back to cached data; show staleness timestampParty size >12Suggest banquet/private dining only; filter out small-table venuesOccasion = "proposal" or "anniversary"Add extra weight to environment + service; flag "special occasion" friendly venues

Error Handling#

Error CodeScenarioHandlingE-SCRAPE-BLOCKEDDianping anti-scraping triggersUse cached data, flag staleness, retry after cooldownE-INSUFFICIENT-REVIEWSRestaurant has <30 reviewsShow with "low confidence" badge; limit recommendationsE-LOCATION-UNKNOWNLocation string can't be resolvedAsk user for district or landmark clarificationE-NO-CANDIDATESZero restaurants match criteriaRelax filters iteratively: expand radius → budget → cuisineE-FAKE-OVERWHELM>60% reviews detected as fakeStrong warning; recommend skipping this restaurantE-PHOTO-PARSE-FAILCan't extract photo metadataDegrade gracefully; skip photo-based fake detection signals

Security Requirements#

• Review data privacy: Scraped reviews used for analysis only; never stored beyond session
• No account impersonation: All scraping is read-only, public-facing data
• Dianping ToS compliance: Respect robots.txt; reasonable rate limiting (1 req/sec)
• Location privacy: City/district level only; never store or expose precise coordinates
• No review manipulation: This tool reads and analyzes; it does NOT post reviews, boost ratings, or interact with Dianping in any write capacity
• Fair use: Analysis results are personal dining decisions; not for commercial competitive intelligence

---

Implementation#

Project Structure#

FilePurposeSKILL.mdFull design document (this file)skill.jsonSkill metadata with script/schema referencesscripts/dianping-guide.pyMain Python CLI — implements all workflow stepsschemas/input.schema.jsonJSON Schema for dining profile and restaurant dataschemas/output.schema.jsonJSON Schema for all output types (authenticity/sentiment/ranking/ordering/route)references/signals.jsonFake review detection signal definitions, weights, and thresholds

CLI Usage#

bash

# Step 1: Define dining context
./scripts/dianping-guide.py profile --location "上海静安寺" --cuisine 川菜 --budget-max 100

# Step 2: Discover candidates (uses built-in sample data)
./scripts/dianping-guide.py discover --cuisine 川菜 --budget-max 100

# Step 3+4+5: Analyze a restaurant (fake detection + sentiment)
./scripts/dianping-guide.py analyze "老王川菜"

# Step 6: Personalized ranking
./scripts/dianping-guide.py rank --cuisine 川菜 --occasion date

# Step 7: Ordering guide
./scripts/dianping-guide.py order "老王川菜" --party-size 2

# Step 8: Exploration route
./scripts/dianping-guide.py route --start "静安寺"

Key Algorithms#

• Fake Review Detection: 8-signal ensemble scoring — time clusters, reviewer age, posting frequency, text similarity (cosine), photo EXIF, generic language, rating patterns, reviewer history. Each signal weighted and normalized to 0-100 authenticity index.
• Sentiment Extraction: Chinese keyword-based matching across 4 dimensions (taste/environment/service/value). Dish-level extraction via pattern matching near sentiment keywords.
• Personalized Ranking: Weighted formula 0.25×cuisine_match + 0.20×value + 0.20×occasion + 0.30×authenticity_bonus.
• Ordering Guide: Separates dishes into Must-order / Recommended / Skip based on aggregated dish-level sentiment from genuine reviews.

Dependencies#

• Python 3.8+ (stdlib only — no external packages required)
• references/signals.json loaded at runtime for detection thresholds

All analysis uses built-in sample restaurant data. For production use, replace SAMPLE_RESTAURANTS with scraped Dianping data (respecting rate limits and robots.txt).