AI Interview Prep: How to Practice Coding, Design and Behavioral Rounds With AI

What AI interview practice is good for, and where it fails

The value of an AI interviewer is volume. Human mocks take money, favors or scheduling, so they are scarce. An assistant gives you a fifth and a fifteenth repetition of explaining an approach aloud, handling a follow-up and recovering from a hint. Grow with Google's interview-tips article recommends AI tools for on-demand practice and names Gemini Live for simulating a spoken interview; OpenAI's ChatGPT Voice page lists interview prep among its uses.

The limits are just as concrete, and the makers state them: Google's help page says Gemini Apps can make mistakes and tells users to double-check responses, and Anthropic's help center says Claude can write things that look correct but are very mistaken. A 2023 research paper found that five AI assistants consistently showed sycophancy, meaning answers that match the user's views over truthful ones, across four text-generation tasks. An interviewer that agrees with you and is sometimes wrong is useful for rehearsal and poor as a judge.

• Grow with Google: interview tips (where the Interview Warmup address now redirects)
• OpenAI: ChatGPT Voice
• Google: Use Gemini Apps (accuracy notice)
• Claude Help Center: incorrect or misleading responses
• Sharma et al. (2023): Towards Understanding Sycophancy in Language Models

Prompt 1: a coding interviewer that does not help too early

Left to its defaults, an assistant tends to explain the optimal solution almost at once, which trains nothing. The prompt below removes that behavior: one problem, no pattern name, no hints unless requested, no code written for you, and a debrief only on request. Fill in the level, role, difficulty and topic, or let it choose the topic so that recognizing the pattern stays part of the exercise.

Run it like a real round. Set a 45-minute timer, state your reasoning before any code, and write in a plain editor without autocomplete if your target interviews use one. Ask clarifying questions first; if the assistant volunteers constraints you did not ask about, tell it to stop. Take the debrief to the verification step that follows, and do not treat its scores as a hiring signal.

You are a senior engineer running a 45-minute coding interview for a <level> <role> position.

Rules:
- Give me one <difficulty> problem on <topic, or "a topic you choose">. Do not name the pattern.
- Wait for my clarifying questions and answer only what I ask.
- Give no hints unless I ask. Then give the smallest hint that unblocks me.
- Never write or fix code for me. If you see a bug, ask a question that leads me to it.
- When I say "done", ask for time and space complexity and for two test cases.
- When I say "debrief": list every bug and missed edge case with an input that triggers it.
  Then score problem solving, coding, testing and communication from 1 to 4,
  with one sentence of evidence for each score.

Start by stating the problem and nothing else.

• Data structures and algorithms interview questions
• Best coding interview prep tools

The verification loop: check the feedback before you believe it

Treat every claim in a debrief as a hypothesis. For code, the check is mechanical: write the slowest obviously correct solution, generate random small inputs, and compare. The script below does that for the classic longest-substring-without-repeating-characters problem. Run with Python 3.12 it printed “2000 random cases agree”. With the condition “and last_seen[ch] >= left” deleted to simulate a bug, the same script stopped at the third case and reported that for the input 'baaacb' it expected 3 and got 5. A counterexample like that is worth more than any review comment.

For design and behavioral feedback there is nothing to execute, so check differently. Ask the assistant to argue against its own feedback and to list what you should verify yourself, then open the documentation for any technical correction before you memorize it. Finally, calibrate against a person at least once: a peer, a colleague or a paid mock interviewer can tell you whether your answer matches the level you are applying for, which a general assistant cannot reliably judge.

• Edge cases: rerun the ones you named aloud, not only the random inputs.
• Complexity claims: time the solution on doubling input sizes and see whether the growth matches.
• Praise: ask what a strict interviewer would mark down, then ask again with “do not soften it”.

import random


def longest_unique_slow(s: str) -> int:
    # Obviously correct: try every substring.
    best = 0
    for i in range(len(s)):
        for j in range(i + 1, len(s) + 1):
            if len(set(s[i:j])) == j - i:
                best = max(best, j - i)
    return best


def longest_unique_fast(s: str) -> int:
    # The sliding-window answer you wrote in the mock.
    last_seen: dict[str, int] = {}
    best = left = 0
    for right, ch in enumerate(s):
        if ch in last_seen and last_seen[ch] >= left:
            left = last_seen[ch] + 1
        last_seen[ch] = right
        best = max(best, right - left + 1)
    return best


random.seed(7)
for case in range(2000):
    s = "".join(random.choice("abc") for _ in range(random.randint(0, 12)))
    want, got = longest_unique_slow(s), longest_unique_fast(s)
    assert want == got, f"case {case}: {s!r} expected {want}, got {got}"
print("2000 random cases agree")

• Sharma et al. (2023): Towards Understanding Sycophancy in Language Models
• Sliding window algorithm patterns in Python

Prompt 2: a system design interviewer that pushes back

Design practice with AI fails in the opposite way from coding: the assistant tends to accept every box you draw. The prompt forces three behaviors a real interviewer has. It withholds requirements until you ask, it questions each component, and it changes a requirement after your first design so you have to adapt instead of reciting. The 80-word cap matters most, because otherwise the assistant delivers the lecture you were supposed to give.

Describe your diagram in words as you go, or paste a text outline of components and arrows. Do the estimates by hand first, let the assistant check them, and settle any disputed figure with a calculator. In the debrief, pay most attention to the list of skipped topics; corrections marked as technically wrong need the same documentation check as before.

Act as a staff engineer interviewing me on system design for 45 minutes.
The prompt is: "Design <system>".

Rules:
- Open with the one-line prompt only. Make me ask for requirements and scale.
- When I state a number, check my arithmetic and tell me if I am off by more than 2x.
- Whenever I add a component, ask what property it provides and what happens when it fails.
- After my high-level design, change one requirement: 10x traffic, a second region,
  or offline clients.
- Keep each reply under 80 words so that I do most of the talking.
- When I say "debrief": list what a strong answer at my level would have covered that I
  skipped, most important first. Mark every statement of mine that you believe is
  technically wrong, and say how confident you are in each correction.

• How to prepare for a system design interview
• 52 system design interview questions

Prompt 3: a behavioral interviewer that asks follow-ups

The risk in behavioral practice is letting the assistant write. A story it drafted sounds fluent and falls apart under follow-ups, because you never lived the details. Anthropic's public guidance for its own candidates draws the same line: use AI to refine your ideas and practice explaining your experience, not to generate experiences you have not had. The prompt therefore gives the assistant your notes and forbids rewriting.

Voice mode suits this round, since the real thing is spoken. Answer aloud, let the assistant ask its two follow-ups, and move on. Afterwards read the debrief for structure only: was there a result, was there a number, which sentence was vague. Whether the story is strong enough for the level is, again, a question for a human.

You are a behavioral interviewer. Below I paste a job description and short notes on
six real stories from my own work.

Rules:
- Ask one "tell me about a time..." question at a time, chosen from the job description.
- After each answer ask two follow-ups that go one level deeper: a number, a decision,
  an alternative I rejected, or what I would do differently.
- Do not rewrite my story and do not invent details. If a sentence was vague, quote it.
- Flag any answer longer than about 300 words; I am aiming for two to three minutes aloud.
- When I say "debrief": for each answer say whether Situation, Task, Action and Result
  were all present, quote my weakest sentence, and say what a skeptical interviewer
  would still doubt.

Job description: <paste>
Story notes: <paste>

• Anthropic: Guidance on candidates' AI usage
• STAR method examples for software engineers

Dedicated AI mock interview tools, as of November 14, 2026

Several interview-prep products now include an AI interviewer next to human or peer sessions. The table records what each company's own site said on November 14, 2026. Features and prices change often, so no prices are listed; check each page before you rely on a feature.

One popular recommendation needs an update. The address long shared for Google's Interview Warmup, grow.google/certificates/interview-warmup, redirected on November 14, 2026 to a Grow with Google article of interview tips, which recommends practicing with Gemini Live. If a page elsewhere calls itself Interview Warmup, check that it belongs to Google before using it.

• interviewing.io
• Aced (formerly Exponent): Practice
• Pramp
• Hello Interview: Guided Practice
• Grow with Google: interview tips (where the Interview Warmup address now redirects)
• OpenAI: ChatGPT Voice

Privacy, and the line between practice and a live interview

Do not paste your employer's code, internal documents or customer data into an assistant to build practice material. Google's Gemini Apps Privacy Hub asks users not to enter confidential information they would not want a human reviewer to see or Google to use to improve its services. Anthropic's Privacy Center explains that consumer chats may be used to improve its models when the user allows it, and that feedback sent with the thumbs buttons is stored. Read the data settings of whichever assistant you use, describe systems in general terms, and change names and numbers in stories.

Practicing with AI is preparation. Using it during an interview is a different act, and companies publish different rules. The safe default is the strictest one: assume no AI in any live round or assessment unless the invitation or your recruiter says otherwise in writing. Covert “interview copilot” tools exist, and Canva's post mentions candidates using them; under a policy like Google's, using one means disqualification.

• Google: Candidate guide for virtual interviews (PDF)
• Anthropic: Guidance on candidates' AI usage
• Canva Engineering Blog: Yes, you can use AI in our interviews (June 11, 2025)
• Google: Gemini Apps Privacy Hub
• Anthropic Privacy Center: Is my data used for model training?
• Is LeetCode dead? What Google's AI interview reports mean

Prompt 4: turn misses into review cards

A mock interview is diagnostic only if the misses come back. After each session, collect what you could not explain: a complexity you got wrong, an edge case, a trade-off you could not name, a number missing from a story. The prompt below turns a debrief into question-and-answer cards and marks the ones the assistant could not verify, so that a hallucinated correction does not become something you memorize.

Review the cards on a schedule and repeat the same mock format a week later to see whether the weakness moved. A spaced-repetition app such as RecallDeck, which publishes this library, can schedule those reviews; the cards support coding practice and mock interviews and do not replace them. For calibration against people, the mock interview platform roundup compares human and peer options.

From this debrief, write review cards for what I got wrong or could not explain.

- One fact or decision per card. Question on the front, answer of at most 40 words on the back.
- Phrase each front the way an interviewer would ask it.
- Skip anything I answered correctly.
- Write UNVERIFIED on any card whose answer you could not check. I will confirm those
  against documentation before I keep them.

Format:
Q: ...
A: ...

• The best mock interview platforms for technical candidates
• How to use spaced repetition for technical interviews

Quick answers

Frequently asked questions