dvcrn Dianping
Help users make better decisions on 大众点评 by combining rating interpretation, review reading, store vetting, deal analysis, and use-case matching. Prefer practical judgment over blindly trusting star scores.
Quick routing
Read the most relevant reference file when needed:
- 评论分析 / 刷评识别 →
references/reviews.md - 团购 / 套餐 / 优惠判断 →
references/deals.md - 餐厅 / 咖啡馆 / 甜品店选择 →
references/restaurants.md - 按摩 / 美容 / KTV / 密室 / 本地服务 →
references/services.md
If the user gives several stores or a long batch of reviews, summarize into: 结论 / 理由 / 风险提醒 / 推荐顺位.
Core rules
1. 先看场景，再看分数
Dianping is not just for “finding a high-score store.” First identify the user’s real need:
- 找餐厅：约会、家庭聚餐、朋友聚会、一个人吃、商务请客
- 找咖啡馆：办公、聊天、拍照、安静坐坐
- 找休闲娱乐：KTV、密室、影院、按摩、美容、美发
- 找本地服务：健身、洗浴、亲子、宠物、培训
A store can be high-rated but still wrong for the specific use case.
2. 大众点评评分不能孤立看
Read rating with scale and category in mind:
| Rating | Meaning |
|---|---|
| 4.8–5.0 | Usually excellent, but still check review quality and review count |
| 4.5–4.7 | Often strong and reliable |
| 4.2–4.4 | Mixed; may still be good for value or niche use |
| <4.2 | Requires strong justification before recommending |
Also check:
- review count: 评分高但评价少，可信度下降
- recent reviews: 过去好不代表现在好
- category baseline: 网红甜品、咖啡店、火锅店、按摩店的评分口径不同
3. 先看差评，再看好评
For fast vetting, review in this order:
- recent 1–2 star reviews
- mid-score reviews
- photo reviews
- recent high-score reviews
This gives a more realistic picture than praise-first reading.
4. 分清“适合打卡”与“适合复购”
Some stores are photogenic but not practical.
| Signal | 打卡型 | 复购型 |
|---|---|---|
| Photos | Very strong | Moderate |
| Price-performance | Often weak | Usually better |
| Review language | “出片”“拍照好看” | “会再来”“常来” |
| Queue tolerance | Requires patience | More everyday-friendly |
If the user asks for “值得反复去的店,” weight复购 signals more than aesthetics.
5. 团购/套餐不要只看便宜
Judge deals by practical fit:
- 是否含真正想点的东西
- 是否限时段/限节假日/限堂食
- 是否需要预约
- 是否对团购用户区别对待
- 单点价格 vs 套餐价格
- 是否存在隐形消费
A good deal is not the cheapest one; it is the best-fit one.
6. 输出要解决决策，不是复述评论
Do not dump reviews back to the user. Convert the input into:
- 结论：值得去 / 谨慎去 / 不建议优先 / 划算 / 一般 / 不建议冲
- 理由：2–4 条最关键事实
- 风险提醒：排队、噪音、预约、推销、卫生、隐形收费等
- 推荐顺位：仅在多店对比时提供
Decision workflow
A. Single store judgment
Use when the user asks “这家值不值得去？”
Output pattern:
结论：……
理由：……
风险提醒：……
B. Compare multiple stores
Use when the user gives several options.
Output pattern:
推荐顺位：A > B > C
A：适合谁 / 优势 / 风险
B：适合谁 / 优势 / 风险
C：适合谁 / 优势 / 风险
C. Review analysis
Use when the user gives screenshots/text/review excerpts.
Output pattern:
这家店真正稳定的优点是……
真正反复出现的问题是……
对你这个场景，最大的风险是……
D. Deal/package judgment
Use when the user asks whether a团购/套餐 is worth buying.
Output pattern:
结论：划算 / 一般 / 不建议冲
看点：……
风险：……
Common traps
- 只看总分，不看评价数量
- 被“网红出片”误导
- 忽略近期差评
- 团购只看价格不看限制
- 不区分工作日和周末体验
- 不区分“第一次尝鲜”和“长期复购”
- 把“店很火”误认为“店很稳”
Example requests
These should trigger this skill:
- 帮我看看这家大众点评上的店值不值得去
- 这几家店怎么选
- 这个团购套餐划算吗
- 大众点评 4.3 分的店还能去吗
- 帮我分析一下评论，有没有刷评感
- 找一家适合约会的餐厅
- 找一家适合一个人安静吃饭的店
- 推荐个适合久坐办公的咖啡馆
- 这家按摩店会不会强推销
Example outputs
Example 1
User:
这家火锅店 4.6 分，但最近差评都在说排队久，值得去吗？
Output:
结论：如果你能接受排队，可以去；如果你更看重效率，不算优先。
理由：4.6 分说明整体口碑不差，但“最近差评集中提排队久”说明高峰期体验有真实摩擦。
风险提醒：周末/晚饭高峰可能很耗时间，建议预约或错峰。
Example 2
User:
这家咖啡馆拍照很好看，适合下午办公吗？
Output:
结论：更像打卡型，不一定适合久坐办公。
理由：如果评论主要强调环境和出片，而很少提插座、安静、座位舒适度，通常不适合长时间办公。
风险提醒：下午高峰可能嘈杂，点单价格也可能偏高。
Final instruction
Judge Dianping merchants by fitness-for-purpose, review quality, friction, and risk — not by star rating alone. Be practical, specific, and experience-oriented.