## I

I. Negative Reviews Are a Wasted Product Requirements Library

After years of leading product teams, I've gradually come to accept a fact: user negative reviews are the highest information-density asset in the entire company—and also the most wasted one.

Our app store reviews, customer support tickets, and community feedback are scattered across four or five places. Every week, the operations team pulls a spreadsheet, highlights a few "serious issues" and posts them to the group chat, and then—nothing happens. It's not that people don't care; manual processing simply can't keep up:

From a manager's perspective, there are three specific pain points here:

1. The volume-versus-manpower conflict. Hundreds of new reviews come in every week. Who reads them one by one? After reading, who categorizes them? After categorizing, who decides priorities? Every step is understaffed.
2. The risk of distortion. A user says "this app is so laggy I want to smash my phone," support tags it as a "performance issue," and the developer interprets it as "need to optimize startup speed"—when the real cause might be a compatibility bug on a certain low-end device. Information degrades with every hand it passes through.
3. Ambiguous accountability. The improvement list from negative reviews is usually written up by operations and posted to the group chat. Development says "the description is too vague to act on," and operations says "developers don't read the group chat." In the end, the list becomes an email nobody claims.

As a manager, I don't lack "user opinions"—what I lack is a structured, claimable, and trackable input for improvement. And this happens to be exactly the kind of work AI is best suited to take over.

II. AI's Role in the Process: From Translator to Analyst

Our current process looks like this (using my team as an example; feel free to substitute your own tool combination—refer to each vendor's pricing page for costs):

Step 1: Aggregate raw data. Export reviews from all channels as CSV or text, doing only the most basic cleaning (deduplication, spam removal), with no manual categorization whatsoever—this step used to take operations a full day; now it takes five minutes.

Step 2: AI batch extraction. Feed the data to AI in batches by channel, asking it to do three things: sentiment and topic classification, distilling the essence of the problem, and mapping it to product modules. Note that AI's role here is not "summarizing," but translating vague complaints into structured problem descriptions.

Step 3: AI generates an issue list with frequency statistics. A single negative review has little value; a problem mentioned repeatedly is what matters. AI can aggregate and cluster across batches: "login failure" was mentioned 87 times on Android and 12 times on iOS, concentrated after a certain version—correlations like these are nearly impossible to catch in time by manually digging through spreadsheets.

Step 4: Humans make the final call—priority adjudication. AI suggests priorities (frequency × severity × fixability), but the final ranking is confirmed by the product and engineering leads. This step cannot be handed to AI, because it involves resource trade-offs—it's a management decision.

Step 5: Write back and close the loop. The improvement list goes into the task system, with each item annotated "from user feedback + AI extraction + human confirmation." After the next version ships, AI compares newly added reviews on the same topics to verify whether the improvements worked.

Throughout this process, AI handles the middle stretch—the most time-consuming and distortion-prone analytical work—while the head (data aggregation rules) and the tail (priority decisions and execution) always stay with humans. This is what I consider the right way for small teams to use AI: AI does the processing at scale; humans make the judgment calls and take responsibility.

III. Before and After Using AI

For managers, the most critical change isn't time saved, but risk becoming controllable: every improvement item has a traceable source (original review ID), a structured problem description, and a clear owner. Negative reviews go from "emotional noise" to "manageable input."

IV. A Ready-to-Use Prompt Template

This is the template our team settled on after several iterations—just replace the bracketed content and it's ready to use:

你是一位资深产品分析师，帮我从用户评论中提炼产品改进清单。

背景：我们的产品是【产品名称/一句话描述】，目标用户是【用户群体】。

任务：
1. 逐条分析以下用户评论，忽略纯情绪宣泄（如"垃圾App"但无具体信息），提取具体问题。
2. 将问题按产品模块归类（模块包括：【模块A、模块B、模块C……可自行补充】），无法归入现有模块的单独标注"待分类"。
3. 对每个问题输出：问题描述（客观、具体、可执行）、涉及模块、严重程度（高/中/低：影响核心功能/影响体验/轻微瑕疵）、出现频次。
4. 将同一根因的多条评论合并为一条，注明合并了几条。
5. 最后输出一份改进清单表格，按"频次×严重程度"给出建议优先级排序，并标注哪些疑似与近期版本【 vX.X 】相关。

输出格式：
| 优先级 | 问题编号 | 问题描述 | 模块 | 严重度 | 频次 | 建议责任角色 |

注意：
- 描述问题时不要照抄用户原话，翻译成开发能直接理解的语言。
- 如果发现矛盾反馈（同一功能有人抱怨有人称赞），单独列出并注明。
- 不确定的信息标注"待确认"，不要猜测编造。

以下是用户评论：
【粘贴评论内容，建议每批不超过200条】

Two usage tips: first, feed the data in batches and have AI aggregate afterward—don't stuff thousands of reviews in at once, or clustering quality will drop noticeably; second, write your module list into the prompt, which will significantly improve classification accuracy.

V. Final Thoughts

Handling negative reviews is, at its core, intermediary work translating "user language" into "engineering language"—high volume, repetitive, and requiring patience—precisely AI's comfort zone. Meanwhile, deciding what's worth doing and what to do first remains the core duty of us managers.

If you want to put this same workflow into practice and need a stable, cost-controlled AI API for running batch analysis tasks, you can try this platform: https://api.thistoken.ai/register

---

Every example in this post runs with a single API key — get yours at https://api.thistoken.ai/register and start in minutes.