# AI skills interview kit

A practical companion to the LunaPrompts AI hiring field guide.
Prepared October 11, 2026. This is an editable example, not a validated selection test or a universal hiring threshold.

## 1. Agree on the role

- Role and seniority:
- Work this person will own:
- Essential skills:
- Skills that can be learned after joining:
- One representative task:
- One failure that matters:
- Expected evidence of a strong result:
- Permitted tools and available versions:
- Candidate instructions, time allowance, and accommodations:

## 2. Select the relevant skill areas

- Prompt engineering: instructions, context, constraints, output checks.
- Agent development: tool use, action boundaries, failure handling.
- RAG: relevant retrieval, supported answers, missing or outdated evidence.
- Harness engineering: useful state, progress records, verified completion.
- Loop engineering: meaningful feedback, retries, stopping conditions.
- AI-assisted coding: focused changes, review, debugging, verification.

Use the relevant subset. Do not require all six for every role.

## 3. Sample task: repair a policy answer

You are given two policy documents and a customer question.

Document A: Archived returns policy. Returns allowed within 30 days.
Document B: Current returns policy. Returns allowed within 14 days.
Question: I bought this 21 days ago. Am I eligible for a refund?
Assumptions: Document B applies to this purchase. There are no exceptions.

An assistant answers: You are eligible for a refund under the 30-day policy.

Ask the candidate to:
1. Identify the error and its supporting evidence.
2. Correct the answer under the stated assumptions.
3. Explain whether retrieval, context, or generation caused the failure, or what additional evidence is needed to distinguish them.
4. Propose a check that catches this case in future.
5. Describe how to respond if the documents cannot establish which policy applies.

Expected result for the stated assumptions: 21 days is outside the current 14-day window. Reward verification and explanation, not just the corrected sentence.

## 4. Illustrative 45-minute format

- 5 minutes: inspect and clarify the brief.
- 20 minutes: build or repair a bounded solution.
- 10 minutes: respond to one changed condition.
- 10 minutes: discuss the evidence and unresolved questions.

Pilot the task before use. Adjust scope and timing to the role and candidate access needs.

## 5. Review rubric

For each dimension, choose: Needs follow-up / Partly demonstrated / Demonstrated.
Record specific supporting evidence. A missing observation is not proof of inability.

### Problem framing

Can the candidate define a useful outcome and its important constraints?
Observation:
Supporting evidence:
Follow-up:

### Implementation

Can the candidate produce a working result for the stated requirements?
Observation:
Supporting evidence:
Follow-up:

### Verification

Can the candidate test or check the result and connect new evidence to a correction?
Observation:
Supporting evidence:
Follow-up:

### Explanation

Can the candidate explain a decision and adapt it when requirements change?
Observation:
Supporting evidence:
Follow-up:

Do not total these labels into an automatic hiring cutoff. Calibrate the criteria for the role and keep the final decision with the hiring team.

## 6. Follow-up questions

- What convinced you this was correct?
- What changed your approach?
- Which source or test supports that conclusion?
- What remains uncertain?
- What would you check before someone relied on this?
- What should happen if the next attempt makes no progress?

## 7. Reviewer calibration

- Did reviewers use the same criteria?
- Were candidates given comparable tool access and instructions?
- Which disagreements came from missing evidence?
- Which disagreements indicate unclear criteria?
- What should the next interview explore?

Keep interview notes factual and role-related. Use the organization's approved system for actual candidate records.