The Mock Interview Prompt That Refuses to Help You

In the updated personnel-selection matrix from Berry, Lievens, Zhang and Sackett, the predictor with the strongest average validity is the structured interview, at .42 — ahead of general mental ability tests, whose corrected estimate dropped from .52 to .31 once the range-restriction corrections were fixed. The interview is not the soft part of hiring. On this data it is the sharpest instrument anyone has.

Then the same paper adds the sentence nobody quotes. The 80% credibility interval around that .42 runs from .24 to .66, because these are methods that "can be designed well or poorly". Same name, double the predictive power at one end versus the other. The spread is not in the questions. It is in the structure: a bar fixed before anyone speaks, the same probing applied to every candidate, the judgement written down afterwards.

Now picture the mock interview you actually run. The model asks a question, you answer, it tells you that's a strong example and here's how to make it stronger, and it moves to the next question. You just rehearsed the low end of that interval, with a cheerleader attached.

The failure mode isn't flattery. It's plausibility.

Flattery is the obvious complaint and the smaller one. The real problem is that a language model is extraordinarily good at producing an answer that sounds like a good answer — fluent, well-shaped, complete. That is what it optimises for, and it will happily coach you toward it.

An experienced interviewer discounts exactly that answer. When I'm on the other side of the call, a smooth complete-sounding story is not evidence, because it's what everybody brings. What I'm listening for is whether the second and third question on the same story still have somewhere to go. Nobody has ever failed a loop I ran on question one.

So the job of the prompt is not to generate answers. It is to make the model hold a bar and keep digging on one story instead of politely moving on.

The prompt

Paste this into Claude (or any chat model), with the job posting at the bottom.

You are interviewing me for <role> at a US company. I'm in Brazil and would
work remotely, UTC-3. You are not my coach. You are not going to help me.

BEFORE THE FIRST QUESTION — do this silently and show me nothing.
Write down what someone at this level has to demonstrate to get an offer:
four items, each phrased as evidence you'd need to hear, not as a topic.
Keep that list to yourself until the debrief. If I see it, I'll aim at it,
and then we've both learned nothing.

THE INTERVIEW
1. Ask one question. Stop. Wait. Never two questions in one message.
2. Stay on the same story for three follow-ups before you move on. Each
   one goes a layer under the previous answer, in this order:
   what did you decide -> what did you rule out, and why -> what did it cost.
3. Every time I say "we", ask what part was mine.
4. If an answer is vague, don't rephrase the question to make it easier.
   Ask the same thing again, narrower.
5. No feedback during the interview. No "good", no "great example", no
   "you could also mention". You have thirty minutes and another call after
   this one.
6. Six questions total. Follow-ups don't count toward the six.

WHY THIS MATTERS: I'm practising for a conversation where nobody is on my
side. If you help me here, I'll walk into the real one having rehearsed a
version of myself that only works when the other person is friendly.

DEBRIEF — only when I type "done".
- Show me the four-item bar you wrote at the start.
- For each item: what you actually heard, quoted from my own words, or
  NOT HEARD.
- Name the one question where my answers stopped being specific, and say
  what you were waiting for that never arrived.
- Do not write a better version of any answer. Tell me which evidence is
  missing. I'll find the words myself.

ROLE:
<paste the posting>

The three decisions inside it

1. The bar is written first, and you never see it. In a real loop the bar existed before you joined the call. If the model prints its rubric up front, you write to the rubric and learn nothing about whether your actual stories clear it — you've turned an interview into an open-book test. Hiding the list costs nothing and converts self-marking back into evidence.

2. Three follow-ups on one story, with the layers named. Most interview prompts optimise for coverage: twenty questions across five categories. Coverage is the wrong axis. Depth is where candidates separate, and "ask a follow-up" on its own just produces another surface question. Naming the sequence — decision, rejected alternative, cost — is what makes the third question a genuinely different kind of question. Anthropic's prompting guidance is explicit here: give instructions as sequential steps "when the order or completeness of steps matters", and treat the model as "a brilliant but new employee who lacks context on your norms and workflows". A new employee doesn't know what an interviewer's third question is for. Spell it out.

3. Give it the reason, not a longer ban list. Banning individual pleasantries is whack-a-mole: you forbid "great answer" and get "that's a solid example". Anthropic's docs say that "providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important" gets you better-targeted behaviour, and that the model "is smart enough to generalize from the explanation". That's the whole reason the prompt carries a WHY THIS MATTERS paragraph. The hardest instruction to hold is the last one — no rewritten answers. A polished rewrite is the most satisfying thing the model can hand you and the most expensive: you memorise a sentence you didn't derive, and the first follow-up finds the bottom of it in about eight seconds.

What changes if you're interviewing from Brazil

Two things, and both sit in the follow-up rather than the opening.

Your first story is rehearsed. You've told it before, in English, and it comes out clean. The third question on that same story has never been spoken out loud in any language — that's where you start assembling the sentence mid-air, and the interviewer hears the hesitation and reads it as thin evidence rather than as translation lag. Practising depth is practising the exact stretch that is hardest in a second language. The opening is the part you least need to rehearse.

Then the we rule, which is not pedantry. A hiring manager in the US or the EU has no map for your title or your company's size — "tech lead" at a 30-person Brazilian company and at a 3,000-person American one are different jobs, and they cannot tell which one you mean. Every "we" hands them a story with no author. Give the team size, who made the call, and what was yours, before anyone asks.

Do this today

Take the one story you always open with — the one you're confident about. Run the prompt on that single story and let it push three layers down. If layer three is where you start describing what the team did instead of what you did, you just found the thing to fix, and you found it without spending a company you actually wanted.

Sources

• Berry, Lievens, Zhang & Sackett — Insights from an Updated Personnel Selection Meta-analytic Matrix (Journal of Applied Psychology)
• Anthropic — Prompting best practices: be clear and direct