Direct answer
Follow-up questions about your project experience go in fixed directions: why you chose it, how you built it, which part was really yours, what metric you measured it by, and why the business needed it.
In one 10-round session we published, the candidate repeated the same summary — “I delivered everything from model integration to the front-end experience” — five times, word for word. The interviewer never changed the subject; it kept drilling into “model integration”: which model, and why that one.
The key is not memorising a smoother answer. It's that for every “I was responsible for X”, you can say at once what exactly you did, why you did it that way, and how you knew it worked.
Editor's note: the follow-up questions on this page come from a published MockLive test session. The candidate's answers were scripted test material voiced by the product's text-to-speech (not a real person); the interviewer's follow-ups are real output of the product's engine. The session was in Chinese, so every quote is shown in the original Chinese with an English translation. The full session and its recordings are in the Chinese article linked in section 4. The author of this article is the author of the product, so this is first-party content; but the five types and the checklist are a general method that neither depends on nor requires our product.
1. Why follow-ups deserve their own preparation
An interview is not a chat. A meta-analysis found that structured interviews — scored on fixed dimensions with a shared rating scale — have a corrected operational validity of about 0.51 for overall job performance, against about 0.38 for unstructured interviews (Schmidt & Hunter, 1998, Psychological Bulletin, 124(2), 262–274, DOI 10.1037/0033-2909.124.2.262; both numbers are meta-analytic estimates of corrected operational validity, not raw observed correlations).
A more conservative re-correction in 2022 did not change the direction: the average operational validity of structured interviews was revised down from 0.51 to 0.42, and of unstructured interviews from 0.38 to 0.19; structured interviews still ranked first among the selection methods studied by highest mean operational validity (Sackett, Zhang, Berry & Lievens, 2022, Journal of Applied Psychology, 107(11), 2040–2068, DOI 10.1037/apl0000994, Table 3 and abstract). Two limits apply: this is an “operational validity” figure that pools predictive and concurrent studies, not a purely predictive coefficient; and ranked by the lower bound of the 80% credibility interval, structured interviews are not first — the paper says so itself.
- Schmidt & Hunter (1998), Psychological Bulletin — DOI
- Sackett et al. (2022), Journal of Applied Psychology — DOI
What this research means for practice is concrete: a structured interview's score comes from asking, dimension by dimension, until there are facts that can be scored. A follow-up question is exactly that move — the interviewer digs down one dimension until they reach a specific detail they can score. Fluency earns nothing; checkable facts do. So what you need to prepare is not a smoother paragraph, but the details underneath every claim.
2. The 5 types of follow-up: real examples and how to answer
These five types summarise the follow-ups that actually came up in this 10-round session; they are not an exhaustive taxonomy — interviewers will of course ask other things. But they cover all ten rounds of this session, and each comes with the engine's original wording so it can be checked against the session transcript.
| Type of follow-up | What it is testing | Real follow-up from this session (original Chinese, with translation) | How to answer |
|---|---|---|---|
| 1 · Why you chose it | Whether you really weighed alternatives, or only memorised a name | 「你能解释一下，在微调大模型时，为什么选择LoRA，而不是其他微调方法？」 — “Can you explain why, when fine-tuning the large model, you chose LoRA rather than other fine-tuning methods?” | Name at least two alternatives that were on the table, what each would have cost, and the specific reason you ruled each out. “We used the mainstream approach” gets another layer of follow-up. |
| 2 · How you built it | Whether you wrote it, or only know the concept | 「你刚才提到后期合成引擎基于FFmpeg自研，能具体说说这个自研方案是如何实现状态恢复的吗？」 — “You mentioned the post-production compositing engine was built in-house on FFmpeg. Can you explain specifically how that in-house solution handles state recovery?” | Give a path down to steps, interfaces or data structures. Describing the feature again is not an answer — the follow-up is asking about exactly the gap between the feature and the implementation. |
| 3 · Where your part ends | Whether “I was responsible” means you did it, or watched it | 「你提到RAG和私有化部署，那具体在一个企业知识库项目里，你个人负责文档解析的步骤吗？为什么是你来主导这一块？」 — “You mentioned RAG and on-premises deployment. In an enterprise knowledge-base project specifically, were you personally responsible for the document-parsing step? Why were you the one leading that part?” | Replace “we” with “I” and draw the line around what you delivered yourself. Saying “a colleague did this part” makes you more believable, not less — clear boundaries are what real work sounds like. |
| 4 · The metric | Whether the result was measured or estimated | 「你提到“前端体验全部交付”，能具体说一下，你是如何衡量和优化用户交互体验的？用过哪些具体指标？」 — “You mentioned you ‘delivered the whole front-end experience’. Can you say specifically how you measured and improved the user interaction experience? Which specific metrics did you use?” | Name the metric you actually watched, how it was defined (how it is calculated), and which change you made moved it and by how much. Give the definition along with the number, or expect “where does that number come from?” |
| 5 · The business | Whether you understand the context, or are reciting | 「你负责合规材料准备中，具体有哪几个关键步骤，是怎样按优先顺序来的？」 — “In the compliance-document preparation you were responsible for, what were the key steps, and how were they prioritised?” | State the order first, and why the first step comes first — that shows you know what matters most in the chain. Skip the order and jump to a conclusion, and you'll be asked to put it in order. |
3. Self-check: run every “I was responsible for X” through this
- Why you chose it — if asked “why A and not B?”, can I name at least two alternatives I evaluated and why I dropped them?
- How you built it — if asked “how exactly did you build it?”, can I go down to steps, interfaces or data structures, rather than describing the feature again?
- Where your part ends — can every “we” in this story become “I”, and can I point to the part I delivered on my own?
- The metric — behind this result, which metric did I actually watch, how is it defined, and how much did it move?
- The business — if the interviewer works in this industry, can I first restate their business context, and then map my experience onto it?
Wherever you get stuck is exactly where a real interview will catch you. Fix that one item first, instead of rehearsing the whole story again — what you rehearse is the narrative; what you fix is the detail.
This checklist is a preparation tool, not a lie detector: published research has not shown that follow-up questions can identify fabricated answers. On the contrary, Levashina & Campion (2007, Journal of Applied Psychology, 92(6), 1638–1656, DOI 10.1037/0021-9010.92.6.1638) found experimentally that follow-up questioning actually increased the degree of faking. So the purpose of these five checks is to fill your detail gaps before you speak — not to hand the interviewer a knife to cut through your story.
4. A transcript of the interviewer following a stock answer all the way down
Below are raw excerpts from that session. The candidate repeated the same summary five times; the interviewer did not change the subject but followed “model integration” down for four rounds in a row. The point of reading it isn't to watch someone fail — it's to see what a follow-up actually looks like: no rebuttal, no change of topic, just one layer deeper each time. And the engine recorded the result faithfully in the scorecard (“mechanically repeated the same stock passage as many as five times”).
| Round | Candidate's answer (transcript, original Chinese — translation) | The follow-up that came next (product output, original Chinese — translation) |
|---|---|---|
| 3 | 「我按项目主线说吧,这几年从算法到全站,我都在一线,最近一个完整闭环是内容生成引擎,从模型接入,路由容错,到前端体验全部交付。您想先深挖哪一段,我展开讲。」 — “Let me go by the main project line: for the past few years I've been hands-on from algorithms to the full stack. The most recent complete loop was a content-generation engine — model integration, routing and fault tolerance, through to the front-end experience, all delivered. Which part would you like to dig into first? I'll expand on it.” | 「你提到“后端模型接入”，具体讲一下是哪个模型？你是怎么接入的？」 — “You mentioned ‘back-end model integration’. Which model specifically? How did you integrate it?” |
| 4 | 「我按项目主线说吧,这几年从算法到全站我都在一线,最近一个完整闭环是内容生成引擎,从模型接入,路由容错到前端体验全部交付,您想先深挖哪一段,我展开讲。」 — (the same summary again) | 「好的，那我们先从模型接入这一块开始，你具体负责采用了什么模型？为什么选这个？」 — “All right, let's start with model integration. Which model were you specifically responsible for adopting? Why that one?” |
| 7 | 「我按项目主线说吧,这几年从算法到全站,我都在一线。最近一个完整闭环是内容生成引擎,从模型接入、路由容错到前端体验全部交付。您想先深挖哪一段,我展开讲。」 — (the same summary again) | 「你能详细说一下里面模型接入这一块，你具体是如何设计和实现的吗？」 — “Can you go into detail on the model-integration part — how exactly did you design and implement it?” |
| 8 | 「我按项目主线说吧,这几年从算法到全站,我都在一线,最近一个完整闭环是内容生成引擎,从模型接入,路由容错到前端体验全部交付,您想先伸拉哪一段,我展开讲。」 — (the same summary again, with one word misheard by speech recognition) | 「你先具体说明一下，在模型接入这一步，你是如何选择和对接当前的大模型的，有哪些具体的技术考量？」 — “First explain specifically: at the model-integration step, how did you choose and connect to the large model you use now, and what were the specific technical considerations?” |
- The full session (in Chinese): the 10-round follow-up chain, the six-dimension scorecard and the replayable recordings
- The report from that session, walked through in English
5. Run it yourself
- Open the practice room, upload your resume (PDF, DOC, DOCX or TXT) and paste the job description you're targeting.
- Pick an interviewer: recruiter / HR screen, hiring manager, technical expert, department head, or executive final round.
- Answer out loud the whole way through; the interviewer follows up round by round on your resume and the job description, and keeps going when you stay vague.
- When it ends you get the six-dimension scorecard: every criticism quotes your own words, with the point in the recording to replay.
The first session is free, no card required; current plans and prices are on the pricing page. The practice room's controls are currently in Chinese — set the interview language to English under 高级设置 (Advanced settings). Scores are model-generated practice feedback, not any employer's hiring signal.
- Product home
- Practise by interview round
- Why you should get stumped here before the real interview
- Pricing and the free session
6. Method and disclosures
The follow-up wording and answer transcripts are taken verbatim from a local end-to-end MockLive test on 2026-09-28 (create session → per-round speech transcription → interviewer engine → scoring). The candidate's answers were scripted Chinese test material voiced by the product's text-to-speech, not a real person's recording; the transcripts contain occasional speech-recognition errors, kept exactly as the system produced them and not corrected by hand. Sample size: 1 session (10 rounds). The five types are the author's summary of this session; they cover every round of it but are neither exhaustive nor a statistical finding about interviewer behaviour, and the “How to answer” column is advice that has not been tested against a control. All three studies were checked against the original sources: Schmidt & Hunter's (1998) 0.51 / 0.38 are in their Table 1 and on p. 267, as corrected operational-validity estimates for overall job performance — not raw observed correlations, and not the training-performance criterion (where interview validity is about 0.35); Sackett et al.'s (2022) 0.42 / 0.19 are in their Table 3, while 0.51 / 0.38 are the earlier values they cite, not their own results; the Levashina & Campion (2007) finding that follow-ups increased faking is taken from the publisher's abstract. The author of this article is the author of the product (first party), not an independent third party; this article is not employment advice.
What this page does not show: that the five types cover every way interviewers follow up; that preparing with the checklist raises your chances of an offer; that follow-ups can expose fabricated answers (the existing evidence points the other way); or that the scorecard agrees with a human interviewer's judgement. Each of these would need repeated independent runs or a controlled study.