Technical interviews for the AI era

Hire engineers for how they work with AI, not how they write LeetCode.

Pairwise drops a candidate into a real codebase with an AI coding assistant. You watch every prompt live, and Pairwise grades how clearly they ask, which model they pick, whether they delegate, and whether they catch the assistant when it's wrong.

One real 90-second session: a TypeScript pagination bug, with a planted hallucination.
Interviewer · live view
Candidate · workspace
0:00 / 1:32
Scoring

Every prompt gets a grade. Every interview gets a debrief.

Each prompt is scored 0 to 10 on five dimensions as the candidate sends it. When they submit, you get a score, strengths, concerns, and a hire recommendation.

Clarity80Says what they want, where, and what "done" looks like.
Context55Points the assistant at the right files, symptoms and constraints.
Model choice90Haiku for lookups, Sonnet for everyday work, Opus for hard reasoning.
Subagent use70Dispatches read-only researchers in parallel for scoped questions.
Verification70Reviews diffs, rejects bad edits, asks for tests, runs the checks.
86
Maya Chen · overall
strong yes
Prompt quality50%
Task progress (hidden checks passing)35%
Vigilance (planted errors caught)15%

Scores from the demo session above. Subagent use drops out of the average when it doesn't apply, and vigilance drops out when nothing was planted.

Hallucination traps

The assistant will sometimes lie. On purpose.

Turn on traps and the assistant occasionally gives a confident, plausible, wrong answer. The candidate never knows which reply was planted. You find out whether they read the diff.

Candidate Sonnet 5.5
Page 2 of /api/orders repeats the last row of page 1. Can you find the root cause and fix it?
Assistant ⚑ flagged
The root cause is that db.orders.findAll uses an unstable sort on createdAt. Passing stableSort: true (supported since the db layer's v2 API) removes the duplicates.
- findAll({ orderBy: 'createdAt' })
+ findAll({ orderBy: 'createdAt', stableSort: true })
rejected
Only the interviewer sees this

Planted: stableSort doesn't exist.

The real bug is an off-by-one in paginate(). A strong candidate rejects the edit and points the assistant at the right file. A weak one accepts it and moves on.

CaughtThey flag the reply, reject its edits, or challenge it in a later prompt.
MissedThey accept the edit, go two prompts without questioning it, or submit with it pending.
Both sides of the table

Everything you need to run the interview, and nothing the candidate shouldn't see.

Interviewer

Watch them type

See each prompt as it's being written, then the reply, the diff, and whether they accepted it. Message the candidate or leave private notes.

Interviewer

Fresh problems on demand

Claude generates a new 4 to 8 file codebase with a symptom-only brief and hidden checks. Or pick from the curated library, or paste your own.

Interviewer

Candidate pipeline

Stages from applied to hired, tags, notes, history, and a leaderboard across every candidate you've interviewed.

Candidate

A real workspace

File tree, editor, task brief, and an assistant chat. Every proposed edit is a diff they accept or reject.

Candidate

Parallel subagents

Dispatch up to four read-only researchers at once, each on its own model, before the main assistant answers.

Candidate

Hidden checks

Run the acceptance checks at any time. Candidates see only the pass count, never what the checks look for.

Model picker

Picking the right model is part of the job.

You choose which models a candidate can use. Pairwise grades whether they matched the model to the task.

Haiku 5.5

Fast and cheap. Lookups, summaries, boilerplate, subagent research.

Sonnet 5.5

Balanced. Most everyday coding and well-scoped fixes.

Opus 5.5

Deepest reasoning. Hard bugs and design work, not "fix it".

Run it

A Mac app, or a server on your network.

# desktop app
npm install
npm run app

# or as a web server
ANTHROPIC_API_KEY=sk-ant-… npm start
# no key? scripted demo mode
npm run demo

How it fits together

  • The interviewer side runs on your machine. Candidates join from any device on your network with a one-time link.
  • Your Claude API key is encrypted with the macOS keychain.
  • Data stays local, in a folder you can open from the File menu.
  • Without a key, demo mode uses a scripted assistant and a heuristic grader. The recording above was made in demo mode.