Hiring guide · for interviewers

AI-Native Engineering Interview Guide

How to interview engineers who build software with AI coding tools (Claude Code, Cursor, Windsurf, Copilot). Nine questions across three tiers, with green and red flags for each, a scoring framework and practical assessments.

Only have 5 minutes? Ask these 3 questions

Each card links to the full question with its green and red flags.

Q1 · Tier 1

“What AI coding tool do you use day-to-day?”

Tests tool awareness and configuration depth.

Q2 · Tier 1

“What percentage of your code do you write manually vs. generate?”

The most honest signal of actual adoption.

Q8 · Tier 2

“Walk me through how you start a new feature with AI.”

Reveals workflow maturity and process thinking.

Scoring framework

Use this to calibrate responses before you start. Most candidates will land at Intermediate. Frontier is rare and worth paying for.

Level	Description	Typical signals
Beginner	Uses AI as autocomplete	Copilot only, writes most code manually, no structured workflow
Intermediate	Active AI user	Cursor/Claude Code, ~50% generated, some prompt engineering
Advanced	AI-native workflow	80%+ generated, custom instructions, verification strategy, sub-agents
Frontier	Builds the harness	Agent harness engineering, framework usage, team process changes, context management

Beginner→Intermediate→Advanced→Frontier

Tier 1: Basics

Any interviewer can ask. These questions require no technical background. A non-technical recruiter or HR manager can ask them and evaluate the responses with the flags below.

Q15-minute screen

“What AI coding tool do you use day-to-day?”

Tests tool awareness and configuration depth.

Green flags

Names a specific tool (Cursor, Claude Code, Windsurf, Copilot) and explains why they chose it, ideally with a comparison to alternatives they've tried.
Describes how they've configured it: custom instructions, CLAUDE.md or rules files, memory features.
Mentions using multiple tools for different tasks.
Can speak to what the tool is good and bad at.

Red flags

Says "ChatGPT" as their primary coding tool. Signals a copy-paste workflow, not AI-native development.
Gives vague answers like "I use AI tools" without naming anything specific.
Has never configured the tool beyond default settings.
Names Copilot but describes using it only for tab completion.

Q25-minute screen

“What percentage of your code do you write manually vs. generate?”

The most honest signal of actual adoption.

Green flags

Advanced candidates say 70–95% generated without hesitation.
Can explain what the remaining percentage is (boilerplate they still type, config files, things where AI reliably hallucinates).
Distinguishes between different types of work: business logic vs. scaffolding vs. tests.
Can explain how the ratio has changed over time and why.

Red flags

Says less than 30% generated. Beginner/Intermediate signal at best.
Gives a number without being able to explain the breakdown.
Says "it depends" and can't give a ballpark.
Seems uncomfortable with the question. AI-native engineers wear this number as a badge.

“How do you verify AI-generated code? What's your quality gate?”

Tests whether verification is systematic or "I read it."

Green flags

Has a systematic approach, not just "I read it."
Describes specific gates: type checker passes, tests cover the behavior, they test edge cases the AI wouldn't think to test.
Explains how they catch AI hallucinations (checking external API references, testing boundary conditions).
Has a mental model for "this output looks right but feels wrong."
Uses CI as a safety net, not a primary gate.

Red flags

Says "I just review the diff" without describing how they review it.
Relies entirely on the AI to self-verify or self-correct.
No mention of tests.
Can't describe a time AI code looked correct but wasn't.
Treats "it compiled" as a quality gate.

Tier 2: Intermediate

Tech-aware interviewer. These questions assume the interviewer has basic technical context. They probe workflow depth and team-level thinking: the difference between a good individual practitioner and someone who can uplift a team.

“Do you have custom instructions, custom skills, or use any kind of framework?”

Tests whether context engineering is treated as a craft.

Green flags

Can describe their setup concretely: a CLAUDE.md or Cursor rules file with specific conventions.
Custom agent skills for recurring tasks, a structured system prompt they maintain and iterate on.
Uses a framework like GSD, Claude.md conventions, or a home-grown approach.
Has opinions on what goes in context vs. what doesn't.
Treats context engineering as a craft.

Red flags

Doesn't know what custom instructions are.
Uses the tool with default settings only.
Has never written a system prompt or rules file.
Thinks prompting is just "telling the AI what you want" rather than a structured discipline.
Can't explain what a CLAUDE.md or .cursorrules file does.

Q85-minute screen

“Walk me through how you start a new feature with AI.”

Reveals workflow maturity and process thinking.

Green flags

Has a repeatable process.
Starts with spec clarity: writing a precise problem definition before touching the AI.
Describes how they load relevant context (which files, architecture docs, existing patterns).
Breaks the feature into smaller tasks for individual agent runs.
Runs verification at each step rather than at the end.
Can describe how the process differs for greenfield vs. editing existing code.

Red flags

Describes an ad-hoc process: "I just describe what I want and see what comes out."
No mention of spec clarity before prompting.
Treats the entire feature as a single prompt rather than decomposing it.
Can't describe what they do differently from how they wrote code without AI.
Describes the AI as doing "most of the thinking." AI-native engineers do the thinking.

“Has your team taken any initiatives to restructure how you work for AI-assisted development?”

Tests awareness that AI changes team dynamics, not just individual productivity.

Green flags

Has concrete examples: changed how tickets are written, moved from Scrum to Kanban, eliminated story points, introduced AI-PR scoring, created shared context files for the team.
Can distinguish between what they changed personally vs. what the team changed.
Has opinions on what's working and what isn't.
Even if their current team hasn't changed, they have opinions on what should change.

Red flags

Says "no, we just use AI for our individual work."
No awareness of how AI changes team dynamics, not just individual productivity.
Hasn't thought about process beyond their personal workflow.
Thinks the only question is "which tool?" rather than "how do we work differently?"

Tier 3: Advanced

Technical interviewer. These questions are for technical interviewers who can evaluate the depth of the answers. They probe frontier-level thinking about agent architecture and context management.

“Do you know what an 'agent harness' is?”

Tests whether they think about the infrastructure that makes AI output reliable, not just the AI.

Green flags

Can define it without prompting: the constraints, CI gates, context files, and verification infrastructure that make agent output reliable. Not the agent itself, but the environment around it.
Has built one or contributed to one.
Can speak to the difference between the agent harness (making individual agents effective) and the human harness (making the team effective with agents).
Has opinions on what makes a good harness.

Red flags

Doesn't know the term.
Confuses the harness with the agent itself.
Thinks "harness" just means prompt engineering.
Has never thought about the infrastructure that makes AI output reliable at scale, only about what the AI does in isolation.

“Do you use sub-agents? How?”

Tests task decomposition and a mental model for parallel work.

Green flags

Uses sub-agents (in Claude Code, Cursor 2.5 cloud agents, or similar) and can explain the task decomposition logic: what tasks get delegated vs. kept in the main context.
Describes how they coordinate output between agents and handle failures in sub-agent runs.
Has a mental model for parallelization: running multiple agents on independent workstreams simultaneously.
Has an opinion on when sub-agents are worth the overhead.

Red flags

Doesn't know what sub-agents are.
Has only ever used a single-agent workflow.
Thinks "sub-agents" just means asking the AI follow-up questions.
Can't describe the coordination problem: what happens when agents work in parallel and produce conflicting outputs.

“How do you handle context window limitations?”

Tests whether they design around the constraint or just hit it.

Green flags

Has concrete strategies: decomposes large tasks into smaller agent runs that fit in context.
Uses structured context files to front-load the most important information.
Knows when to start a fresh context rather than trying to cram everything in.
Knows what information belongs in the harness (persistent) vs. in the prompt (per-task).
Has experience debugging hallucinations caused by context drift, where the model "forgets" earlier parts of a long conversation.

Red flags

Doesn't know what a context window is.
Treats it as a fundamental blocker rather than a manageable constraint.
Has never designed around it. Just hits the limit and starts over.
Thinks bigger context windows make the problem go away (they don't; they just move the bottleneck).

What level should you expect?

Use this table to calibrate expectations before the interview. Frontier candidates are rare. Don't hire an Advanced person for a role that needs Frontier, and don't pass on an Intermediate because you were hoping for Advanced.

Role	Minimum level	Ideal level	Key signal
Junior	Beginner	Intermediate	Learning fast, asks good questions about AI tools
Mid	Intermediate	Advanced	Consistent AI workflow, some custom tooling
Senior / Lead	Advanced	Frontier	Verification strategy, harness thinking
Staff+	Frontier	Frontier	Team process changes, opinions on restructuring

Beyond the questions: practical assessments

These three formats complement the interview questions and reveal things that verbal answers can't.

45 min

Live debugging session

Provide a small, intentionally broken codebase. Watch how they prompt, verify, and handle AI hallucination, not the fix itself.

30 min

Portfolio walkthrough

Ask the candidate to screenshare a recent project and walk through one feature from prompt to production. "Where did the AI get it wrong? What context did you provide?"

20 min

Agent design sketch

Present a simple product requirement, ask them to decompose it into agents or sub-agents. No code, just evaluating their mental model.

Skip LeetCode. It evaluates skills that agents now handle.

Red flag quick reference

Dealbreakers

Claims to use AI tools but cannot name specific tools, workflows, or failure modes
Cannot describe any verification process for AI-generated code
Philosophical resistance to AI-generated code in a role that explicitly requires it

Serious concerns

Only uses AI for autocomplete or boilerplate
No awareness of context window limits or prompt degradation
Has never encountered a case where AI-generated code failed

Minor / contextual

Unfamiliarity with specific agent frameworks by name (they change fast)
Hasn't restructured a team around AI (may not have had the opportunity)
Uses a different toolchain than your company uses (tool transfer is easy, mindset transfer isn't)

A note on fairness

AI-native engineering has only been real since late 2025. Nobody has even a year of experience. Questions about tool adoption can correlate with factors beyond skill: company policy, team culture, access. Focus on reasoning quality and learning trajectory, not specific tool names. Ask "even if you haven't had the authority to restructure a team, how would you approach it?" to open questions to candidates at any career stage.

For deeper context on what AI-native engineering actually is, beyond tools and into the mental model shift, read up on the Bottleneck Flip (generation is cheap; specification and verification are the constraint) and the two harnesses (the agent harness that makes one agent reliable, and the human harness that makes a team effective with agents).