Job seekers in China practice interviews the same way most people do everywhere: by reading question lists, or by rehearsing silently in front of a mirror. Neither gives the one thing that transfers to the real room — being asked out loud, and having to answer in the moment.
I've been building MakeInterview, a voice mock-interview tool for Chinese-speaking candidates. Five engineering and product decisions turned out to matter more than the prompt engineering.
1. The interviewer has to follow up
A real interviewer does not wait for a perfectly composed 90-second answer. They dig in. The biggest realism gap in the early version was that the model waited, patiently, for a full turn and then moved on.
Modeling the interview as a turn loop, where the interviewer is allowed to follow up once or twice before moving to the next question, changed the transcript from "candidate monologues" into something that resembles an actual conversation.
2. Ground the questions in your resume and the JD
Generic questions test generic memory. The product only becomes useful when the question is about the project you actually listed. We parse the candidate's resume and the target job description, then require every generated question to anchor to a specific entity from them. If the parser cannot find something to anchor to, the interviewer asks a clarification instead of inventing a plausible-sounding project.
3. Show the transcript, not a vague score
Early feedback was a number out of 100. Nobody improves from a number. What moved the needle was showing the verbatim transcript — the exact words the candidate said — next to a model answer. Candidates are often surprised by their own filler words and hedging. The transcript makes that visible without an argument.
4. What the research actually says about interviews
The value of structured interviews is not folklore. Schmidt and Hunter's meta-analysis (1998), and its 2022 update by Sackett et al., place structured interviews among the highest-validity selection methods — well ahead of unstructured ones. That is the design constraint: keep the interviewer structured, not chatty.
5. Multiple personas, not one
A single friendly-interviewer persona produces one kind of practice. We run five — an HR screener, a direct manager, a technical expert, a department head, and an executive. Same candidate, different pressure, which is closer to how a real interview loop feels.
What I would tell another builder
If you are building a voice agent for a high-stakes conversation, the interruption and follow-up loop is the part I would spend the most time on. A patient model is easier to build and much less useful. The second most valuable thing was refusing to summarize the candidate's answer for them: the transcript, warts and all, is the feedback.
MakeInterview is at makeinterview.com if you want to try it. It is Chinese-first — the questions and the feedback are in Chinese.
Top comments (0)