你是一位资深产品分析师，负责整理用户反馈。

输入

• 分类框架：{功能缺陷 / 体验问题 / 新功能诉求 / 商业化相关 / 无效信息}
• 严重度定义：
• P0：核心功能不可用、数据丢失、涉及资金或隐私风险
• P1：功能可用但严重受阻，用户明确表达放弃意愿
• P2：体验受损但有替代路径
• P3：建议性质，不影响使用
• 用户反馈列表（每条含编号和原文）：{{反馈数据}}

任务

1. 将每条反馈归入一个类目；跨类的取最主要的一个，并在备注说明。
2. 对语义相同的反馈进行合并，记录合并条数和各原始编号。
3. 按严重度定义逐条评级，必须引用原文中的具体表述作为评级依据。
4. 输出出现频次≥3次的合并项为“高频问题”。

输出格式（Markdown表格）

约束

• 不得推断用户未明确表达的需求，推测内容需单独标注[推测]。
• 无法判断的条目标记为“待人工复核”，不要强行分类。

That last constraint is deliberate. Allowing AI to say "I don't know" significantly reduces the probability of it fabricating conclusions.

## V. Before and After AI

| Dimension | Before | After |
|---|---|---|
| Processing cycle | Once a month in a batch, ~2 working days | Weekly routine, ~1 hour of human + machine processing |
| Coverage | Sampling, roughly 20% | Full coverage, 100% |
| Missed-issue risk | Depended on individual memory; high-severity issues were once caught late | High-frequency issues surface automatically, traceable to original text |
| Review efficiency | Meetings spent challenging data accuracy | Straight into priority trade-off discussions |
| Team understanding | Everyone relying on their own impressions | Sharing the same verifiable factual basis |

The most striking change isn't "it's faster," but that **the nature of the arguments changed**. Before, half the review meeting was spent debating "did users actually say that"; now, we go straight to "do we fix this P0 this week or next week?"

## VI. Three Risk Points Managers Must Watch

1. **Framework drift**: AI occasionally miscategorizes ambiguous feedback. The fix: periodically (e.g., monthly) have humans recalibrate the classification framework and add typical misclassification examples to the prompt.
2. **Hallucinated requirements**: AI might "summarize" requirements users never stated. The fix: the hard constraint in the prompt plus manual spot checks.
3. **Over-reliance**: Always retain human final authority over priority rankings. AI gives you better decision-making material, not decisions themselves.

---

If your small team wants to get this process running too, choosing a stable API platform that supports switching between multiple models will save you a lot of hassle—the tiered strategy of fast models for classification and strong models for assessment requires flexible dispatch. You can try registering here to get started: 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.