/

/

How to hire prompt engineers

How to hire prompt engineers

To hire a prompt engineer, test prompts against cases they haven't seen: live prompt writing, a real assistant to fix with AI tools and a follow-up on hidden tests.

Prompt engineer loop4 rounds · 3 AI interviewers
Amara · Prompt engineerShortlisted
  1. 1✓AI screening interviewA prompt she took from demo to productionAriaReport ready
  2. 2✓Live prompt writingSupport emails in, strict JSON outMaxReport ready
  3. 3✓Build and testA support assistant, scored on 60 hidden casesNixReport ready
  4. 4✓Follow-upHer prompt, her tests and the assistant itselfNixReport ready
Your team decidesOne report per round, with the evidence

The prompt engineer interview process

Four rounds, each with its own report. Every claim is captured and scored against your rubric.

Aria · Live interview
Aria

Tell me about a prompt you took from demo to production.

Amara"A claims-intake assistant. It invented policy numbers about one time in thirty. I rebuilt the prompt around a 300-case test set, and got that down to one in five hundred."
Aria

How did you know the fix held up, and wasn't just tuned to your test set?

Amara"I kept a separate set the prompt never saw, and checked live traffic every week."

Round 1 · AI screening interview

Prompts they took from demo to production

Aria starts with a real conversation and follows up on the answers. Every claim is captured and scored against your rubric.

Nix · Build and test
TaskHere's a freight company's support assistant and its refund policy. Make it follow the policy, resist manipulation and pick the cheapest model that holds up. Show your tests.
Cursor
Tyler

Keeps asking for stricter rules. The prompt balloons, and he tests one message by hand.

Claude Code
Amara

Has it draft tests, adds the tricky ones it missed and fixes the problem they catch.

Same assistant. Same models. How they used the tool made the difference.

Round 3 · Vibe coding interview

A real assistant, scored on cases they never see

You choose the models and the budget. Candidates work in their own tools, and Nix scores the finished assistant on 60 hidden test cases.

NINIX · Follow-up interview20 min
NI

Your batch stalled partway through. How do you resume without duplicates?

C

Answering · 01:42

Max · Live prompt writing
JSON
1{
2  "order_number": "…",
3  "order_numbers": ["…", "…"],
4  "issue_type": "…",
5  "urgency": "…"
6}
MaxWrite a prompt that pulls the order number, issue type and urgency from a support email, as strict JSON. Talk me through it.Amara"I'll give it the schema, one short example, and tell it to answer with JSON only."MaxWhat happens when an email mentions two orders?Amara"Right now, it picks the first one. I'll make order numbers a list, and add an example with two."

Round 2 · Live prompt writing

Live prompt writing against real test emails

No AI assistant helping. Just the candidate, real test emails and an interviewer who asks questions as they type. This round checks the fundamentals.

Nix · Follow-up
NixHidden case 17. The customer wrote in Spanish, and the assistant answered in English. The policy says reply in their language. Why didn't it?Amara"My examples were all English, so the model copied them. I'd add a Spanish example, and a language test."NixYou switched to the smaller model. What would make you switch back?Amara"If refund-policy cases drop below a hundred percent, or escalations rise in production."

Round 4 · Follow-up interview

Hidden test cases and model choices

Nix opens the candidate's prompt, their tests and the assistant itself, then asks about the cases it got wrong. Nix scores how they work with AI, not just whether the demo looks good.

Aria · Live interview
Aria

Tell me about a prompt you took from demo to production.

Amara"A claims-intake assistant. It invented policy numbers about one time in thirty. I rebuilt the prompt around a 300-case test set, and got that down to one in five hundred."
Aria

How did you know the fix held up, and wasn't just tuned to your test set?

Amara"I kept a separate set the prompt never saw, and checked live traffic every week."

Round 1 · AI screening interview

Prompts they took from demo to production

Aria starts with a real conversation and follows up on the answers. Every claim is captured and scored against your rubric.

Max · Live prompt writing
JSON
1{
2  "order_number": "…",
3  "order_numbers": ["…", "…"],
4  "issue_type": "…",
5  "urgency": "…"
6}
MaxWrite a prompt that pulls the order number, issue type and urgency from a support email, as strict JSON. Talk me through it.Amara"I'll give it the schema, one short example, and tell it to answer with JSON only."MaxWhat happens when an email mentions two orders?Amara"Right now, it picks the first one. I'll make order numbers a list, and add an example with two."

Round 2 · Live prompt writing

Live prompt writing against real test emails

No AI assistant helping. Just the candidate, real test emails and an interviewer who asks questions as they type. This round checks the fundamentals.

Nix · Build and test
TaskHere's a freight company's support assistant and its refund policy. Make it follow the policy, resist manipulation and pick the cheapest model that holds up. Show your tests.
Cursor
Tyler

Keeps asking for stricter rules. The prompt balloons, and he tests one message by hand.

Claude Code
Amara

Has it draft tests, adds the tricky ones it missed and fixes the problem they catch.

Same assistant. Same models. How they used the tool made the difference.

Round 3 · Vibe coding interview

A real assistant, scored on cases they never see

You choose the models and the budget. Candidates work in their own tools, and Nix scores the finished assistant on 60 hidden test cases.

Nix · Follow-up
NixHidden case 17. The customer wrote in Spanish, and the assistant answered in English. The policy says reply in their language. Why didn't it?Amara"My examples were all English, so the model copied them. I'd add a Spanish example, and a language test."NixYou switched to the smaller model. What would make you switch back?Amara"If refund-policy cases drop below a hundred percent, or escalations rise in production."

Round 4 · Follow-up interview

Hidden test cases and model choices

Nix opens the candidate's prompt, their tests and the assistant itself, then asks about the cases it got wrong. Nix scores how they work with AI, not just whether the demo looks good.

NINIX · Follow-up interview20 min
NI

Your batch stalled partway through. How do you resume without duplicates?

C

Answering · 01:42

Aria · Live interview
Aria

Tell me about a prompt you took from demo to production.

Amara"A claims-intake assistant. It invented policy numbers about one time in thirty. I rebuilt the prompt around a 300-case test set, and got that down to one in five hundred."
Aria

How did you know the fix held up, and wasn't just tuned to your test set?

Amara"I kept a separate set the prompt never saw, and checked live traffic every week."

Round 1 · AI screening interview

Prompts they took from demo to production

Aria starts with a real conversation and follows up on the answers. Every claim is captured and scored against your rubric.

Max · Live prompt writing
JSON
1{
2  "order_number": "…",
3  "order_numbers": ["…", "…"],
4  "issue_type": "…",
5  "urgency": "…"
6}
MaxWrite a prompt that pulls the order number, issue type and urgency from a support email, as strict JSON. Talk me through it.Amara"I'll give it the schema, one short example, and tell it to answer with JSON only."MaxWhat happens when an email mentions two orders?Amara"Right now, it picks the first one. I'll make order numbers a list, and add an example with two."

Round 2 · Live prompt writing

Live prompt writing against real test emails

No AI assistant helping. Just the candidate, real test emails and an interviewer who asks questions as they type. This round checks the fundamentals.

Nix · Build and test
TaskHere's a freight company's support assistant and its refund policy. Make it follow the policy, resist manipulation and pick the cheapest model that holds up. Show your tests.
Cursor
Tyler

Keeps asking for stricter rules. The prompt balloons, and he tests one message by hand.

Claude Code
Amara

Has it draft tests, adds the tricky ones it missed and fixes the problem they catch.

Same assistant. Same models. How they used the tool made the difference.

Round 3 · Vibe coding interview

A real assistant, scored on cases they never see

You choose the models and the budget. Candidates work in their own tools, and Nix scores the finished assistant on 60 hidden test cases.

Nix · Follow-up
NixHidden case 17. The customer wrote in Spanish, and the assistant answered in English. The policy says reply in their language. Why didn't it?Amara"My examples were all English, so the model copied them. I'd add a Spanish example, and a language test."NixYou switched to the smaller model. What would make you switch back?Amara"If refund-policy cases drop below a hundred percent, or escalations rise in production."

Round 4 · Follow-up interview

Hidden test cases and model choices

Nix opens the candidate's prompt, their tests and the assistant itself, then asks about the cases it got wrong. Nix scores how they work with AI, not just whether the demo looks good.

NINIX · Follow-up interview20 min
NI

Your batch stalled partway through. How do you resume without duplicates?

C

Answering · 01:42

Prompt engineer skills to assess

Score every candidate on the same six skills, rated against evidence from the rounds.

  • Prompt writing
  • Structured output
  • Testing prompts
  • Handling tricky inputs
  • Choosing models
  • Working with AI tools
Scorecard · Prompt engineerRisk: Low
Prompt writingVery strong
Structured outputStrong
Testing promptsVery strong
Handling tricky inputsStrong
Choosing modelsVery strong
Working with AI toolsVery strong
Criteria met6 of 6

Prompt engineering interview questions

Use them in your own loop, or let Aria, Max and Nix ask them, follow up on the answers and score them.

1

Tell me about a prompt you took from demo to production.

2

How did you know the fix held up, and wasn't just tuned to your test set?

3

Write a prompt that pulls the order number, issue type and urgency from a support email, as strict JSON.

4

What happens when an email mentions two orders?

5

Make a support assistant follow the refund policy, resist manipulation and use the cheapest model that holds up.

5

The customer wrote in Spanish and the assistant answered in English. Why?

5

You switched to a smaller model. What would make you switch back?

Integrity in every round

Bad Actor Detection flags calls to AI models outside the allowed list, a second voice or screen, and work pasted from outside the workspace.

Hidden test cases

Every assistant is scored on cases the candidate never sees, so one lucky run doesn't pass.

We evaluate. You decide.

Hello Recruiter evaluates and organizes the evidence. Your hiring team decides who moves forward.

Frequently asked questions

Still have a question?

See the loop on your own roles.

How do you hire and interview prompt engineers?

Test prompts against cases they haven't seen. Hello Recruiter runs a live prompt-writing round with test inputs and structured output, a real task in the candidate's own AI tools, and a follow-up on the system prompt, a hidden test case and a model comparison.

What should a prompt engineering test include?

Realistic inputs, a required output format such as JSON, hidden test cases the candidate didn't tune for, and a comparison across models, so results reflect how well the prompt holds up rather than one lucky run.

How does Hello Recruiter detect cheating or misrepresentation?

Every round is monitored for integrity signals such as calls to AI models outside the allowed list, a second voice or a second screen, and answers or work pasted from outside the workspace. Each signal is timestamped in the candidate's integrity report, and claims are cross-checked across rounds.

What does the hiring team get after each round?

A report for every round: an AI score out of 100, evaluation criteria marked Met, Not met or Not yet evaluated, skills rated Very strong, Strong, Weak or Very weak, a risk level, and any fraud signals with timestamps. Hello Recruiter evaluates; the hiring team decides.

More AI hiring guides

How to hire AI engineers

Test how they build, review and ship with AI.

Vibe codingFollow-upAI system design

How to hire AI product managers

Test judgment on AI quality under pressure.

Take-home specVibe prototypingRole-play

All hiring guides

Every role in the series

See all guides

Run this loop on your next prompt engineer hire

Precision hiring, with the evidence behind every pick.