Opening with a Counterexample: I Treated AI as an "Answer Machine" and Wasted a Week

Last year, I wanted to build a bookkeeping tool for freelancers. With no budget for real user interviews, I turned to AI: I opened a chat window and asked directly:

> "What bookkeeping features do freelancers need most?"

The AI gave me a beautiful feature list: invoice management, multi-currency support, tax reminders, report visualization… every item looked reasonable. I spent a week building according to that list, showed it to a few potential users, and got this feedback: "All these features already exist in other software."

Where did it go wrong? I treated the AI as a "requirements answer generator." What it gave me was the "average consensus" of freelance bookkeeping software on the internet, not the real pain points of my target users. I made three common mistakes:

1. Letting the AI answer out of thin air — without any user personas or scenario constraints, it could only output generic knowledge;
2. Asking "what features are needed" instead of "what problems are encountered," skipping the motivation layer;
3. Treating AI output as a conclusion rather than a hypothesis to be validated.

This is the most common failure mode: using AI to simulate interviews is, in essence, just looking up an "industry encyclopedia" somewhere else.

The Right Approach: AI Is a "User Role-Player," Not a "Requirements Consultant"

After I turned the corner, I redesigned the process. The core mindset shift: AI shouldn't answer "what users want" — it should play the role of a specific user so I can interview it. What AI can do for me:

• Play multiple differentiated personas: Generate 5-8 virtual users at once with different backgrounds, personalities, and usage habits, covering edge-case users you might never encounter in real interviews;
• Refuse to play along with my assumptions: Explicitly instruct it in the prompt to "not just agree with my product," simulating the pickiness and indifference of real users;
• Output structured interview records: After each simulated conversation, automatically organize the notes into a three-column table of pain points, motivations, and alternative solutions, making hypothesis extraction easy;
• Ask "why" endlessly without embarrassment: In real interviews it's awkward to repeatedly probe "why," but with AI you can dig all the way down.

The Process I Finally Got Working

Step 1: Define a set of user personas. First, have the AI generate a batch of candidate personas based on my product direction. I manually filter out the obviously implausible ones and keep 5 differentiated ones (e.g., "a recently resigned illustrator" and "an employed programmer taking freelance side gigs" have completely different pain points).

Step 2: Conduct simulated interviews one by one. Activate only one persona at a time, and ask open-ended questions using a semi-structured outline: When was the last time you dealt with bookkeeping? How did you do it then? What frustrated you? Note: don't mention my product yet.

Step 3: Extract hypotheses, not conclusions. After the interview, have the AI compress the records into several falsifiable hypotheses, e.g., "The core reason freelancers procrastinate on bookkeeping isn't laziness — it's that their income and expense channels are too fragmented."

Step 4: Validate at low cost. Take these hypotheses to real communities, post about them, and talk to 5 real people for 20 minutes each. Simulated interviews can't replace real validation, but they turn my real validation from "aimless wandering" into "confirming or refuting hypotheses" — the hit rate is completely different.

Before and After Comparison

A real interview costs several hundred yuan per session and takes a week to schedule; a simulated interview round takes half an hour and can be re-run anytime. It's not perfect — AI-played users may be overly cooperative or overly typical — but as a hypothesis generator and a testing ground for interview outlines, its cost-effectiveness is unmatched.

A Reusable Prompt Template

Save the template below, replace the bracketed content, and it's ready to use:

你是一位用户访谈模拟器。请扮演一位用户，接受我的一对一访谈。

【用户画像设定】
- 身份背景：[如：30岁，自由插画师，工作2年]
- 使用场景：[如：多个平台接单，收入来源分散]
- 性格倾向：[如：对财务工具不耐烦，习惯拖延]
- 现有替代方案：[如：Excel + 微信收藏夹]

【扮演规则】
1. 始终保持该画像的视角，不要跳出角色给建议；
2. 不要主动提及或猜测我的产品，除非我先提到；
3. 如果我问诱导性问题（如"你是不是觉得X很麻烦"），
   只有在符合你设定时才同意，否则给出真实反应（包括"没想过""不在乎"）；
4. 回答保持口语化，长度不超过3句话，可以反问我；
5. 你的痛点要具体到事件和时间，不要说空泛的"效率低"。

【输出约定】
当我说"访谈结束"时，你再切换成分析员角色，
把本次对话整理成表格：用户原话 | 背后动机 | 对应需求假设（可证伪）。

现在请用一句该用户会说的开场白开始。

A practical tip: keep each interview to no more than 15 rounds. When you're done asking, say "访谈结束" first to get the structured analysis, then start a new session with the next persona.

A Few Practical Tips

• Make personas a bit "extreme." Too-average personas only produce too-average insights. Deliberately setting up difficult users (e.g., "someone who has tried three bookkeeping apps and abandoned them all") is more valuable;
• Ask the same question across 3 personas. If three personas give similar pain points, that hypothesis's priority can be moved up;
• Model choice: Role-playing and resistance to leading questions are sensitive to model quality. Use a stronger model for the simulation stage; a cheaper model is fine for the analysis stage. As for API costs, refer to each platform's official pricing page;
• API key management: If you want to turn this process into a scheduled batch job, remember to route through a server-side proxy and never hardcode your key into a script repository — but that's another story.

Simulated interviews aren't a replacement for real interviews — they're the step that comes before: using extremely low cost to rehearse "what should I ask, and whether my hypotheses hold up." What indie developers and small teams lack most is a trial-and-error budget, and what this step saves you is directional waste.

If you're planning to run this process, you'll need a stable AI API to support batch role-playing and structured analysis. You can try this platform: https://api.thistoken.ai/register

---

Tired of juggling provider integrations? Register at https://api.thistoken.ai/register and call every model through one base_url.