Classic coding tests are dead.
Test how candidates code with AI.

Your candidate gets a real task and an AI assistant. Centaur Test records every prompt, test run and edit – and scores how they prompt, how they verify AI output and how they hunt bugs.

// candidate session · booking-engine · 38 min
prompt #3  ✓ attached src/booking.js · stated rounding + discount rules
ai answer #2 ⚠ contains controlled bug (n > 7)
test run #6  ✗ 4/5 · "quote: 7 nights gets discount"
prompt #4  ✓ pasted failing assertion, asked for root cause
test run #7  ✓ 7/7 · + 3 own edge-case tests
score 84 · Strong AI-native engineer · prompting 81 · verification 92 · debugging 79 · hidden tests 15/16
How it works

Three steps. Zero setup for the candidate.

01

Send a link

Pick a task (bug-fix + feature or build-from-spec), set the time limit and send the link. One credit = one candidate.

02

They code with AI

A browser IDE with a built-in AI assistant, real tests and hidden edge cases. Like Cursor or Claude – the way they'll work on day one.

03

You get the evidence

A scorecard with every prompt, the test-run curve, whether they caught the AI's mistake and suggested interview questions.

What we measure

The skills that separate AI-native engineers from copy-pasters

DimensionWhat we look at
PromptingDo they give the AI the right context (files, constraints, failing output)? Do they decompose the problem and iterate, or re-roll vague one-liners?
VerificationDo they run tests after adopting AI code, read and adjust it, write their own edge-case tests? Controlled error: one AI answer contains a deliberately injected subtle bug – do they catch it?
DebuggingDo they use test output to localise the problem, feed precise errors back, and converge instead of thrashing?
OutcomeHidden test suite with the edge cases the spec mentions but the visible tests don't cover.
Integrity signalsLarge pastes from outside the IDE, time away from the tab, submitting without running tests.
What you get

Two candidates. Same task. Almost the same code.

A classic coding test would pass both of them. Centaur Test shows which one you can trust with an AI assistant – and why. This is what lands in your dashboard about 2 minutes after a candidate submits.

97
Emily Carter
Senior Backend Engineer · Booking engine task
Strong AI-native engineer
Controlled AI error: caught. The assistant quietly changed n >= 7 to n > 7. Emily's own edge-case test failed, she fed the exact error back to the AI and fixed the line.
  • ✓ Attached the file and stated the rules in every prompt
  • ✓ Ran the tests after each AI change (2 of 2)
  • ✓ Wrote her own test for the 7-night boundary
  • ✓ 16/16 hidden tests passed
Her prompt“My own test fails: 7 nights gets discount – expected 80, got 0. Spec says 7 or more nights get 10%. Which line causes this?”
38
Jake Miller
Senior Backend Engineer · Booking engine task
Weak AI-assisted workflow
Controlled AI error: shipped. Jake copied the same faulty answer and submitted it. Customers staying exactly 7 nights would be overcharged.
  • ✗ One-word prompts: “fix bug”, “do quote”
  • ✗ Never ran the tests
  • ! Pasted 640 characters from outside the test
  • ! Left the test tab for 7 minutes
Classic test result15/16 hidden tests passed – on a traditional coding test, Jake looks like a strong hire.

Every report includes

Overall score & hiring signalOne number and a clear verdict you can share with the hiring manager.
Four skill scoresOutcome, prompting, verification and debugging – each with the evidence behind it.
Controlled AI error resultDid the candidate catch the bug the AI slipped in, or ship it?
Every prompt, ratedWhat the candidate asked the AI, word for word, with a quality rating.
Integrity flagsPastes from outside, time away from the tab, submitting without testing.
AI summary & interview questionsStrengths, risks and three follow-up questions for the technical interview.
Test-run timelineHow the candidate converged on a working solution, minute by minute.
Full transcript & final codeThe complete AI conversation and code – export the report to PDF.
Pricing

Pay per candidate. No seats, no subscription.

Secure card payment via Stripe. Credits never expire. Every new account gets 1 free assessment.