I'm in the middle of a job change right now, here in Japan. Mid-career, aiming for a presales role in the cloud industry. And while preparing for interviews, a slightly meta question occurred to me: everyone says "the job market has changed because of AI" — so why not ask the AI itself how to win?
So I did. I had a long conversation with Claude about what actually matters in Japanese job interviews now that AI is everywhere. Some of the answers were what I expected. A few were not. This post is the distilled version, plus a small AWS project that came out of it: a mock interviewer built on Amazon Bedrock.
A caveat before we start: this is written from a Japanese perspective. Job hunting in Japan has its own rituals — the 職務経歴書 (a detailed career history document), the almost-mandatory question about why you're leaving your current company, the 逆質問 ("do you have any questions for us?") that is secretly part of the evaluation. If you're interviewing elsewhere, the principles probably still transfer, but the examples are local.
What AI actually changed in Japanese hiring
The first thing the AI pointed out was blunt: your resume no longer differentiates you.
Two years ago, a well-structured 職務経歴書 with quantified achievements put you ahead of most applicants. Today, anyone can paste their work history into ChatGPT or Claude and get a polished document in five minutes. Recruiters know this. Some Japanese companies now run AI screening on incoming resumes too, so you have AI-written documents being filtered by AI readers.
The result is not that documents don't matter — a sloppy one still disqualifies you. It's that a good one has become table stakes. The differentiation has moved into the interview room, where it's still just you, talking.
The second change is a new standard question. In my own interviews this year, and in what I hear from other candidates, some version of this keeps coming up:
「普段、AIをどう使っていますか？」 — "How do you use AI in your daily work?"
This question didn't exist in 2023. Now it functions the way "do you know Excel?" did twenty years ago, except the bar for a good answer is much fuzzier. Which brings us to the AI's actual advice.
The advice, distilled into three moves
Move 1: Bring proof, not polish
Since everyone's documents read well now, the AI's suggestion was to shift effort from polishing words to producing artifacts. A public GitHub repository, a deployed app, a blog post, a conference talk — things AI cannot fake on your behalf, because they exist outside the interview.
This matched my own experience more than anything else in the conversation. I've built a dozen or so small apps over the past year, most of them on AWS, and pushed them to GitHub with live URLs. In interviews, "I'm good at catching up with new technology" is a claim. "Here's a URL, it's running right now, and here's the repo" ends the discussion. Japanese interviewers in particular tend to probe claims gently and repeatedly (「具体的には？」— "concretely?"), and an artifact short-circuits that entire loop.
Move 2: Prepare your "how I use AI" story with numbers
For the new standard question, the AI's advice was to prepare it like you'd prepare 転職理由: a short, concrete story with a before and after. Not "I use ChatGPT for research" — that's the 2026 equivalent of "I can use a computer."
A good answer has the shape: what task, what tool, what changed, and — importantly — what you still check by hand. That last part is what separates a thoughtful answer from a hype answer. Japanese companies are noticeably cautious about AI-generated mistakes reaching customers, so saying "AI drafts it, and here is exactly what I verify before it goes out" lands better than pure enthusiasm.
Move 3: Rehearse against an AI interviewer, not a mirror
The classic Japanese interview prep is writing out 想定問答 (anticipated Q&A) and rehearsing alone. The AI's point was that this practices recall, not conversation. What trips people up in real interviews is the follow-up question — the second and third 「なぜ？」 that you didn't script.
An LLM is genuinely good at this. It can play a specific interviewer persona, push back on vague answers, and do it at 11pm the night before your interview without anyone's patience running out. This is the part I decided to actually build, since I had the pieces already.
Building a mock interviewer on Amazon Bedrock
If you're an AWS user, you don't need a SaaS subscription for this. The Bedrock Converse API and a persona prompt get you a working mock interviewer in about twenty lines of Python:
import boto3
client = boto3.client("bedrock-runtime", region_name="us-east-1")
SYSTEM = """You are a hiring manager at a mid-sized Japanese IT company,
interviewing a mid-career candidate. Conduct the interview in Japanese.
Ask ONE question at a time. Cover, in order:
1. 転職理由 (reason for changing jobs)
2. A strength, then drill into it with follow-up questions until
the answer is concrete (names, numbers, what the candidate did personally)
3. How the candidate uses AI in daily work, including what they verify by hand
If an answer is vague, do not move on — ask 「具体的には？」 and dig deeper.
After the final question, give frank feedback in Japanese on each answer."""
messages = []
print("面接を始めます。終了するには exit と入力してください。\n")
while True:
user_input = input("あなた> ")
if user_input.strip() == "exit":
break
messages.append({"role": "user", "content": [{"text": user_input}]})
response = client.converse(
modelId="global.anthropic.claude-sonnet-4-6",
system=[{"text": SYSTEM}],
messages=messages,
)
reply = response["output"]["message"]["content"][0]["text"]
messages.append(response["output"]["message"])
print(f"\n面接官> {reply}\n")
A few notes from actually running this:
- The instruction "do not move on if the answer is vague" is the whole product. Without it, the model politely accepts weak answers, which defeats the purpose
- Asking for frank feedback at the end works better than asking for feedback after every answer, which turns the session into a lecture
- Change the persona and you change the interview. "A skeptical engineering manager who dislikes buzzwords" produces a completely different — and more useful — hour than a generic HR persona
Model availability differs by AWS account, so check what you can invoke with the Converse API first; any recent Claude model on Bedrock handles this fine. If you want to go further, the same loop drops nicely into a small web app, but honestly the terminal version is enough for interview prep.
One thing that surprised me: the model's feedback on my 転職理由 was harsher than any human mock interviewer I've had. Humans soften criticism, especially in Japan. The AI just told me my second paragraph sounded like a complaint about my current employer. It was right, and I rewrote it.
What the AI said it can't do for you
I asked the obvious closing question: "So can you get me hired?" The answer was a clear no, for reasons worth repeating.
First, Japanese hiring still weighs 人柄 — roughly, character and how you'll fit into the team — and that is evaluated live, in the room, through unscripted interaction. No amount of preparation output survives contact with a genuinely curious interviewer if there's nothing real behind it.
Second, and this should go without saying: using AI to feed you answers during a live interview is a terrible idea. Beyond the ethics, the latency and the unnatural phrasing are detectable, and in Japan's relationship-driven hiring culture, being caught would end not just that application but possibly your standing with the recruiting agent too.
The honest framing the AI gave me was this: it can compress your preparation time and raise your floor, but the ceiling is still yours. The resume it polishes, everyone has. The rehearsal it provides, anyone can do. What's left as differentiation is exactly the old stuff — real experience, real artifacts, and being a person the interviewer wants to work with.
Which is a strangely reassuring answer to get from the machine that caused the problem.
Closing
If you're job hunting in Japan right now, the practical takeaways are short: treat your documents as a pass/fail gate rather than a weapon, prepare a concrete "how I use AI" story with a verification step in it, put at least one artifact where an interviewer can see it, and rehearse against an AI that's been told not to be nice to you.
Total cost of my Bedrock mock interviewer so far: a few dozen yen. Cheaper than one coaching session, and it never gets tired of my 転職理由.
Good luck out there.
Top comments (0)