You can read forty interview question lists and still freeze on the first follow-up. The fix is a mock interview you run on a schedule, with a score sheet you fill in afterwards, so practice produces numbers instead of vibes. This post gives you the mock interview script for AI engineer roles: a 60-minute run order, a weighted score sheet, and a prompt that turns a chat model into an interviewer that does not just compliment you.
Everything here is a practice design I wrote. The prompts are untested templates: I have not run them against a live model for this post, so treat the first run as a calibration run. For question content, use AI engineer interview questions on RAG, agents and evals and the system design walkthrough. This post is the harness around them.
What does a good mock interview script for AI engineers include?
Three things that most practice sessions skip. A fixed run order, so you rehearse stamina and transitions. A rubric with observable behaviours, so a score of 3 means the same thing every time. And a record, so you can compare run 1 to run 3.
The 60-minute order below mirrors a common shape of AI engineer loops (a project discussion, a design problem, a short technical probe), but I am not claiming any specific company runs it. Adjust it to the job post you actually have.
| Minutes | Block | What the interviewer does |
|---|---|---|
| 0 to 5 | Intro | Asks for a 90-second summary of you and one project |
| 5 to 20 | Project deep dive | Probes one project: decisions, numbers, failures |
| 20 to 45 | Design problem | Gives a one-line brief (for example a document Q&A bot) and lets you drive |
| 45 to 55 | Technical probes | Four or five short questions on evals, cost, latency, retrieval |
| 55 to 60 | Your questions | You ask; interviewer notes quality of questions |
The score sheet
Score each row from 1 to 5 right after the session, before you look at anything else. The anchors are the part that makes it repeatable: you score the behaviour, not your feelings.
| Row | Weight | 1 means | 3 means | 5 means |
|---|---|---|---|---|
| Clarifying questions | 10 | Started designing at once | Asked about users and scale | Asked what a wrong answer costs and set a quality bar before designing |
| Structure | 10 | Wandered | Had blocks but lost time | Stated a plan, kept to it, and adapted when redirected |
| Technical correctness | 20 | Wrong or hand-wavy claims | Mostly right, some gaps | Correct, with trade-offs named |
| Evaluation thinking | 20 | "I would check manually" | Named metrics | Described a labelled set, how to split retrieval from generation errors, and how to catch regressions |
| Numbers | 10 | None | Stated, some unchecked | Stated as assumptions, multiplied correctly, sanity-checked |
| Failure modes | 10 | None | Listed generic ones | Named a specific failure, how you would detect it, and the fix |
| Ownership in the project talk | 10 | "We" throughout | Mixed | Clear on what you decided and what you would change |
| Communication | 10 | Rambling, jargon | Clear in places | Concise, checked in with the interviewer, said "I don't know, here is how I'd find out" when true |
The weights sum to 100. Weighted score is the sum of each row's score times weight, divided by 5. A quick worked example with made-up scores: if you score 4 on a weight-20 row, that contributes 4 x 20 / 5 = 16 points. All fives gives 100, all threes gives 60, all ones gives 20.
Do not read too much into the total. The weights are my judgment, not a published hiring standard, and no employer uses this exact sheet. The value is the lowest two rows. Those are your next three days of practice.
The mock interviewer prompt
Paste this into a fresh chat. It is an untested template, so run it once on a throwaway answer first.
You are a senior engineer interviewing me for an AI engineer role.
Job post (paste below). My project (paste a 150-word summary below).
Run a 60-minute mock interview in this exact order and wait for my
answer after every message:
1. Ask me to introduce myself and my project in 90 seconds.
2. Project deep dive: ask ONE question at a time. After each answer
ask one follow-up about the weakest part (a missing number, an
unjustified decision, "what did you personally do").
3. Design problem: give me a one-line brief based on the job post.
Answer clarifying questions briefly and in character. Do not
give hints unless I say "hint".
4. Five short technical probes on evaluation, cost, latency and
retrieval or tool use.
5. Ask "What questions do you have for me?"
Rules: no feedback until I type SCORECARD. Do not invent facts about
the company. If I make a technical claim you believe is wrong, ask a
question that exposes it instead of correcting me.
On SCORECARD: score me 1 to 5 on these rows: clarifying questions,
structure, technical correctness, evaluation thinking, numbers,
failure modes, ownership, communication. For every score, quote the
exact sentence of mine that justifies it. A 5 requires the full
behaviour in the anchor. If you are unsure between two scores, give
the lower one. Then list the two weakest rows and one drill for each.
[PASTE the anchor table from the score sheet here]
Why "give the lower one" and "quote the exact sentence"? Chat models lean generous, and a score with no quote behind it can't be checked. The second rule also lets you spot when the model scored a sentence you never said.
Calibrate the scorer before you trust it
Take a deliberately poor answer, such as "I would use a vector database and GPT, and check if it works", paste it as your design answer, and type SCORECARD. Evaluation thinking and failure modes should come back at 1 or 2. If they come back at 3 or above, tighten the anchors, or switch to a different model for scoring than the one you used to interview. Using one model for both is convenient and makes bias easier to miss.
How to run the session so it counts
- Time yourself out loud. Typing hides pacing problems. If you can, use voice mode or speak to a recording and transcribe it, then paste the transcript. Voice features vary by app and plan, so check what yours supports.
- No notes in the project block. You will not have your README open in the real one. If you cannot remember your own chunk size or your eval result, that is a finding. The project README template exists partly so you know your own numbers.
- Fill the sheet within ten minutes. Memory of what you said fades fast.
- Keep a log. One line per run: date, total, lowest two rows, the drill you chose.
| Run | Date | Weighted total | Lowest rows | Drill chosen |
|---|---|---|---|---|
| 1 | ||||
| 2 | ||||
| 3 |
Leave it blank in your copy and fill it in. A rising total matters less than the lowest rows changing.
Drills for the rows people lose most
- Evaluation thinking. Take one project. Write 20 questions with expected answers, then compute retrieval hit rate separately from answer correctness. The method is in building evaluation datasets. Then say aloud: "If the number dropped, here is how I would tell retriever from generator."
- Numbers. Redo a cost estimate from scratch with different assumptions. The cost per 1000 requests post has worked arithmetic you can copy the method from.
- Failure modes. Write three real failures of your own project, each with symptom, detection and fix. If you do not have three, break your project on purpose and record what happens.
- Ownership. Rewrite your project summary replacing every "we" with "I decided" or "my teammate decided". It feels awkward and interviewers notice the difference. The guide to explaining a RAG project in an interview gives a structure for the project block.
Using a human for one run
A model cannot see your face, and it cannot tell that you rushed the last ten minutes. Ask a friend or a peer to read your score sheet and sit through one run, even if they are not in AI. Give them the anchors and ask them to circle any row where your answer was hard to follow. A non-expert marking "I lost you here" is useful data on communication, the one row models score least reliably.
If you have a take-home or live coding round, this script does not cover it. For employer rules on AI use during hiring, check the specific company's candidate guidance. Anthropic, for example, publishes guidance saying AI is encouraged for interview preparation but not allowed in live interviews unless told otherwise (page read on 2026-10-09, showing "last updated Jul 10, 2025"; confirm current wording on the page itself). Other employers differ, so ask your recruiter.
A one-week plan using the sheet
- Day 1: Run 1, fill the sheet, pick the two lowest rows.
- Days 2 to 3: Drill those two rows on your own project only.
- Day 4: Run 2 with the same job post. Compare rows, not totals.
- Days 5 to 6: Fix whatever moved the wrong way, and rehearse your questions for the interviewer.
- Day 7: Light run with a human, then rest.
For prompts that tailor questions to your CV and job description, the earlier mock interviewer prompt post covers behavioural rounds. This one is built for the technical loop.
