## A Manager's Old Problem: When Requirements Get Finaliz...

A Manager's Old Problem: When Requirements Get Finalized, Whose Voice Is at the Table

Anyone who has led a product team knows this scene: in a sprint planning meeting, everyone argues over a feature for two hours, and in the end, whoever is loudest wins. After the meeting, someone mutters under their breath, "Do users really need this?" — but nobody can answer, because arranging a single real user interview — from screening and outreach to summarizing notes — usually takes one to two weeks, and the deadline is right around the corner.

This is exactly my entry point: AI-simulated user interviews don't replace real research; instead, they transform "refining requirement hypotheses" from something dependent on one individual's intuition into a process step that the whole team can observe, trace, and iterate on repeatedly. For independent developers and small teams short on people and budget, it's often the only realistic option.

First, the Pain Point: It's Not That We Don't Want Interviews — It's That the Cost Structure Doesn't Hold Up

If you break down user interviewing, the costs fall into three areas:

First, recruiting is hard. Small teams have no user pool. In the cold-start phase, you can't even scrape together ten target users willing to chat for half an hour — and you still have to offer red envelopes, schedule times, and deal with no-shows.

Second, information loss is severe. Even when an interview happens, the notes depend on one person scribbling furiously. The interviewer's questioning style, depth of follow-ups, and even their mood that day all affect the conclusions. The same interview, heard by two people, produces two different directions.

Third, conclusions can't be audited. That line in the final PRD — "users want Excel export support" — nobody knows which sentence it came from, in what context, or whether it was the consensus of three users or one person's grumbling. When managers review requirements, they're actually reviewing secondhand information that can't be verified.

The combined result of these three: requirement decisions in small teams live in a permanent state of "gut call plus after-the-fact justification," with all the risk pushed to post-launch.

What AI Can Do: Turn Interviews from an "Event" into a "Process"

I've recently been rolling out a practice in small teams: before any real interviews, run a round of AI-simulated interviews first to produce a structured list of requirement hypotheses; then real interviews (even just three) are used to calibrate those hypotheses. AI plays the role of a "hypothesis generator." The process has four steps:

Step 1: Define the persona. The team writes out together who the target user is: role, use scenario, current workaround, biggest complaint. Don't skip this step — a vague persona produces a useless simulated interview. I recommend co-creating it in a team collaboration doc, with the manager signing off.

Step 2: Run the simulated interview. Feed the persona to the AI, have it play this user, and let team members take turns asking questions. The key technique is to require the AI to "have emotions, hold opinions, and be allowed to refuse to answer" — not a perfect interviewee who answers everything. A simulated user who never pushes back has no value.

Step 3: Cross-cutting synthesis. Have the AI produce structured output based on multiple conversation rounds: a pain-point list, requirement hypotheses ranked by frequency and intensity, and the original dialogue excerpts corresponding to each hypothesis. The API service I use charges by token, so a round of simulated interviews costs almost nothing (see the pricing page on the official site for details) — anyone on the team can spin one up on a whim.

Step 4: Hypothesis triage and validation scheduling. Classify hypotheses into "high-risk, high-uncertainty" and "low-risk, can be validated later." The former go into the real interview outline; the latter go straight into the release plan and get validated with data after launch.

After running this process, what the manager receives changes: no longer "I think users want X," but a hypothesis list where every item is annotated with its source, confidence level, and validation method. In review meetings, the subject of discussion shifts from opinions to evidence chains.

Before and After: Same Three Weeks, Completely Different Deliverables

Before the transformation, here's what one round of requirement research actually looked like: we scheduled 4 users, 2 canceled last minute; the product assistant compiled the notes, and 10,000 words of transcripts were condensed into 5 requirements with no basis for prioritization; during sprint planning, whether to build one item — "users want offline mode" — sparked a 90-minute argument, finally settled by vote.

After the transformation: we first ran simulated interviews with 6 different personas (two days, including team members rotating as interviewers), producing 17 requirement hypotheses, 5 of which were flagged as highly uncertain. We scheduled only 3 real users, but the entire outline centered on those 5 hypotheses — every session went straight to the point in 30 minutes. In the final PRD, every requirement could be traced to two layers of evidence: "simulated hypothesis + real validation." Review time dropped from two hours to 40 minutes.

What's saved isn't just time — it's risk control. During the simulated interview phase, the team discovered that two seemingly obvious requirement hypotheses didn't hold up against AI-played extreme users, and cut them outright, avoiding about two weeks of development work. This ability to "falsify early" is the missing piece in traditional processes.

A Reusable Prompt Template

Below is the system prompt we use for simulated interviews. You can copy it directly and swap out the persona section:

你将扮演一位目标用户，配合我完成一场模拟访谈。请严格保持角色，不要跳出。

## 用户画像
- 身份角色：{例如：30人电商团队的一名运营主管}
- 使用场景：{例如：每周手动整理各平台销售数据做周报}
- 当前替代方案：{例如：Excel + 复制粘贴，耗时约4小时}
- 最大抱怨：{例如：平台口径不一致，对账总出错}
- 性格与表达习惯：{例如：直接、不耐烦、只关心省时间}

## 扮演要求
1. 有立场：对自己的工作方式有偏好，被质疑时会反驳或解释
2. 有边界：与画像无关的问题可以说"这个我不清楚"或"没想过"
3. 有情绪：如果问题踩到痛点，表现出明显的抱怨倾向并展开说
4. 拒绝完美：不要主动总结需求，不要替产品说话，只描述自己的经历

## 访谈结束后
当我说"访谈结束"时，切换为分析师角色，输出：
1. 痛点清单（按提及频率和情绪强度排序）
2. 需求假设（每条附原文依据片段）
3. 高不确定、需要真实访谈验证的假设（不超过5条）

One usage tip: have the person on your team closest to the users write the persona section — the ceiling of the AI's role-play quality is determined here. After a round, paste the AI's hypothesis list back into your team doc, collect contrasting opinions in the comments, and run the next iteration.

A Few Reminders

Simulated interviews cannot replace real users — that's the bottom line. The AI-played user comes from the persona you provide; if the persona is biased, the conclusions inherit the bias. So the correct framing is: AI makes hypotheses explicit and structured, real interviews calibrate them, and data delivers the final verdict. Only with these three layers of evidence stacked together can requirement decisions be called controllable.

For teams ready to try this, choosing a stable model API service is a prerequisite — simulated interviews need to run repeatedly, across many rounds, so interface stability and cost predictability matter more than single-shot capability. The service I currently use supports switching between multiple models, so you can use different models for simulated interviews and structured synthesis. Registration link here: 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.