原始反馈没法直接拍动作。从 Intercom、App Store、Typeform、Linear 导出的 500 条原话,把真正的信号埋在噪音底下。下面 15 个 Prompt 把这些反馈聚类、计数、打标签、排优先级,变成 roadmap 真正能回答的一小撮主题。每个模板都覆盖支持工单、NPS 评论、流失退订调研、App 评论、beta 转录——粘到 ChatGPT、Claude 或 Gemini,再替换方括号里的占位符即可。
速览
- 先用模板 1(开放分类法),等你摸清类别后再用模板 2 重跑。
- 永远要求每个主题给计数——没数就没法排优先级。
- 任何聚类结果给到高管前,先跑模板 15(幻觉守门)。
- 100-500 条短回复,任何前沿模型都够用。整季度(一次几千条原话)就用 1M token 的模型:Claude Sonnet 4.6、Gemini 3.1 Pro,或 GPT-5.5(应用内完整 1M 仅 ChatGPT Pro $200 档才有)。
- 粘贴前先去掉 PII(姓名、邮箱、账户 ID)——这些 Prompt 默认输入是干净的。
选哪个模型、一次能喂多少
聚类 Prompt 的成败,取决于一次能塞进多少原始反馈。瓶颈已经不是模型质量,而是上下文管理。截至 2026 年 6 月,应这样选:
| 模型 | 上下文窗口 | 聚类强项 | 备注 |
|---|---|---|---|
| Claude Sonnet 4.6 | 1M token | 总结紧凑、幻觉低、原话还原好 | 主力机型;含在 Claude Pro $20/月 内 |
| Claude Opus 4.7 | 1M token | 处理模糊、混杂反馈的推理最强 | 偏贵(API $5/$25 每百万 token);留给难的 JTBD / 严重度判断 |
| Gemini 3.1 Pro | 1M token | 最便宜的 1M 前沿模型(API $2/$12) | 单次大批量强;含在 Google AI Pro $19.99/月 |
| GPT-5.5 | 1M(API) | 稳妥默认;选择器有 Instant/Thinking/Pro | Plus $20 档应用内约 320 页,完整 1M 仅 Pro $200/月 |
实用经验:约 750 个英文词≈1,000 token,所以 1M 窗口能装约 70 万词原话,远超任何单季度的量。真正的失败模式是一次问太多、得到一堆浅主题,而不是撞到 token 上限。如果一批结果聚得很泛,就按周或按 segment 分批,再对 summary 聚一次。
要做可复用、全团队规模化的整合,专用平台(Dovetail、Thematic)能补上原话可追溯和与 NPS/CSAT 挂钩的情绪打分。Prompt 胜在速度和灵活,平台胜在可审计。当不止一个团队要查询同一份反馈语料、且每个主题都要追溯到源原话时,就该考虑上专用平台。
适合哪些人
负责季度反馈整合的 PM 和 CX 负责人、自己读支持 inbox 的创始人、跑大调研的 research ops,以及在找下一个实验切入点的增长团队。
什么时候别用这些 Prompt
n 小于 30 不要用——直接手读每一条。反馈高度专业(工程日志、code review)也别用——模型缺这套领域分类法。还有,绝不要把原始 PII 喂给模型:先清掉姓名、邮箱、账户 ID。
Prompt 结构拆解
一个反馈聚类 Prompt 应带这六个要素:
- 角色:让 AI 扮演谁(资深 PM / 独立创始人 / 产品设计师 / 增长负责人)。
- 上下文:阶段(想法 / MVP / 增长 / 规模化)、团队规模、流量或 ARR、平台(web / iOS / Android)、受众、限制。
- 目标:一个具体交付物——一组主题、一批 ticket、一份变化对比。
- 限制:主题数量、互斥性、严重度量表、必须从源文本引用。
- 输出格式:表格、矩阵、可贴 ticket 的 JSON,或带标签的段落,能直接粘进 Linear / Notion / Jira。
- 源校验:用来引用的原始原话,外加一条”标出任何不在源文本里的引用”的指令。
15 个可直接复制的 Prompt 模板
[N]、[paste] 这类方括号是占位符——替换成你的数字和粘贴的反馈。
1. 首轮主题提取
从这里开始。先让 AI 自己提出分类法,再去约束它。
You are a senior PM clustering raw user feedback. Below are [N] verbatim responses. Group them into 8-12 themes. For each theme: (1) short label (3-5 words), (2) one-sentence description, (3) count of responses, (4) 2 representative verbatim quotes, (5) suggested severity (blocker / major / minor / nice-to-have).
Feedback: [paste]
可替换变量: N、feedback corpus。
优化建议: 主题重叠时追加:“Merge themes where 60%+ of underlying quotes could fit either. Keep clusters mutually exclusive.”
2. 约束分类法聚类
已经知道你在乎哪些类别时用。
Cluster the feedback below into these predefined categories: [bug, missing feature, pricing, onboarding, performance, integration, support quality, other]. For each: count, % of total, top 3 verbatim, single recommended action. Anything in "other" exceeding 10% should suggest a new category.
[paste]
3. 情绪 + 主题矩阵
Cluster the feedback below into themes AND label each with sentiment (positive / mixed / negative). Output a matrix: theme by sentiment, with counts. Highlight any theme where the same feature gets equal positive and negative comments — that is where the actual disagreement lives.
[paste]
4. 流失原因聚类
Below are [N] cancellation-flow survey responses. Cluster the reasons for leaving into 6-8 categories. For each: count, %, 2 verbatim, whether it is reversible (product fix vs life-circumstance), and one suggested counter-move.
[paste]
5. NPS 评论按分段聚类
Cluster these NPS comments into themes, but split the output by score band: detractors (0-6), passives (7-8), promoters (9-10). For each band: top 5 themes with counts. Highlight any theme that appears in BOTH detractors and promoters — that is your polarizing feature.
[paste]
6. Bug 与功能请求分流
Take this mixed feedback corpus and split it into two stacks: bugs (something broken vs expected behavior) and feature requests (something not built yet). For each stack, cluster into themes with counts. Flag any items where the line is unclear.
[paste]
7. 按 persona 聚类
Below is feedback tagged with user persona (free / paid / enterprise / new / power). Cluster themes BY persona — the same theme can appear in multiple personas. Output: theme by persona matrix with counts. Highlight which themes are concentrated in paid + power users — those move revenue.
[paste]
8. JTBD 重构
Re-cluster this feedback using JTBD framing instead of feature categories. Output 5-8 jobs in the form "When X, I want Y, so I can Z." For each: count of underlying feedback, the products / workarounds users currently use for that job, and where our product falls short.
[paste]
9. 严重度 × 频率 2x2
For each theme in this clustered feedback, place it on a severity (low/high) by frequency (low/high) 2x2. Output as a table. The high/high quadrant is the next sprint. The high-severity / low-frequency quadrant needs an audit (rare but bad).
[paste themes + counts]
10. 每主题挑代表原话
For each of these themes, pick the 3 most representative verbatim quotes for an internal share-out. Criteria: clarity, emotion, specificity. Exclude any quote with PII (names, emails, account IDs). Tag each quote with persona if known.
Themes + raw quotes: [paste]
11. 跨渠道交叉核对
I have 3 sources of feedback for the same quarter: app reviews, support tickets, sales loss reasons. Compare the top 5 themes from each. Output a table: theme by source. Highlight themes appearing in all 3 sources (highest confidence) and themes appearing in only 1 (channel-specific noise or hidden signal).
[paste 3 source summaries]
12. 每主题对应动作
For each theme below, propose 1 product action, 1 GTM / messaging action, and 1 thing NOT to do. Each action should be testable in 2 weeks. Mark which action belongs to which team.
[paste themes]
13. 季度变化对比
Compare last quarter's feedback themes to this quarter's. Output: themes that grew, themes that shrank, new themes, retired themes. Hypothesize 1 reason for each major delta. End with the 3 themes worth a deep-dive next quarter.
Last quarter: [paste]
This quarter: [paste]
14. 反馈转 ticket
Convert the top 5 themes from this clustering into engineering / design tickets. Each ticket: title (under 12 words), problem (2 sentences from clustered evidence), proposed scope, success metric, linked verbatim quotes. Output as JSON for Linear / Jira import.
[paste themes + quotes]
15. 幻觉守门轮
唯一不可省的模板。任何聚类结果送到决策者手里之前,先跑它。
Audit this AI-generated clustering against the source feedback. For each theme: confirm the count by spot-checking, flag any verbatim quote that does not appear in source, flag any claimed pattern not supported by at least 3 quotes. Output: confirmed themes vs themes to redo.
Clustering: [paste]
Source: [paste 20 random verbatim]
容易踩的坑
- 让 AI 出 25 个主题,其实 10 个就够——那是在过拟合噪音。
- 每个主题不给计数——没数就没法排优先级。
- 让模型编造原话——一定要给原文,并要求对照核查(模板 15)。
- Bug 和功能请求混进同一聚类——动作和团队都不同。
- 忽视情绪——同一功能 50/50 正负面,正是你最有争议的点。
- 不和上季度对比就重聚——漂移本身就是信号。
- 少于 5 条原话就开搞——太薄,不值一个 sprint。
怎么把结果再推进一步
- 粘贴前先去掉 PII:姓名、邮箱、账户 ID。
- 任何聚类结果往上汇报前,先跑模板 15。
- 每个主题至少配 3 条原话;一条只是个例。
- 每季度跑一次模板 11,把真信号和渠道噪音分开。
- 从源工具(Intercom、Linear、Typeform)重新导出,别复用旧导出——主题每周都在变。
- 聚类计数都很均匀时要怀疑;真实信号通常是 2-3 个主题占主导。
- 分享前给每个主题指定一个 owner,否则没有动作落地。
FAQ
- AI 一次能聚多少条?: 任何前沿模型一次干净处理 100-500 条短回复都没问题。整季度(几千条原话)就用 1M token 的模型——Claude Sonnet 4.6、Gemini 3.1 Pro 或 GPT-5.5——大约能装 70 万词。再多就按周或 segment 分批,再对 summary 聚一次。
- 哪个模型最适合?: 日常聚类,Claude Sonnet 4.6 总结紧凑、幻觉低。最便宜的大批量单次调用是 Gemini 3.1 Pro(API $2/$12 每百万 token)。最难的 JTBD 或严重度推理用 Claude Opus 4.7。均为 2026 年 6 月数据。
- 怎么确认 AI 没在编?: 跑模板 15,并随机抽 10 条原话回源比对。若有 2 条是编的,整批重做。
- 原话直接喂给 AI 吗?: 是,但先去 PII。原话是最高价值的输入,转述会丢掉信号。
- 多久聚一次?: 活跃产品每月一次,最少每季度。周期之间的漂移本身就是领先指标。
- 主题太泛怎么办?: 加一条约束:主题 label 必须引用一个功能、屏幕或工作流——禁用”体验""质量”这类抽象词。
相关阅读
标签: #Prompt #产品创业 #User Story