Take-homes

Take-homes where AI is welcome.

A real task from your codebase, for engineering roles. Then a call about their own code.

bookingOrrery take-home, 90 minutesRecordingHand it in
src
booking.ts
slots.ts
tests
TASK.md
AI this session
Claude Code14 lines
Claude Code wrote these 2 linesPrompt: "hold the slot, release it if confirm fails"
Claude Code
Larpy joinedRight after submit
CandidateIf confirm throws, finally still releases it. In a catch it would leak on success.
Larpy, on line 13
Take-home result
Backend engineer
0
Your team decides
Your tests
10 of 10
Hidden edge cases
4 of 5
Understood own code
3 of 3
Written with Claude Code
14 lines
“If confirm throws, finally still releases it. In a catch it would leak on success.”Code call, 2:14, on line 13
18:42
Q1Q2Q3
Code call, answers pinned to linesWatch the full recording
Handed in. 10 of 10 tests pass. Larpy is joining.

Set it up in a chat.

Point Larpy at a merged pull request. It drafts the task, the tests and the answer key.

Acme Payments / Backend engineer, payments / Coding assessmentDraft, savedTurn on for this job
Context
acme/payments-apiPR #2281 Retry failed payoutsJob description
Make the assessment from the payout retry PR. Mid-level, 90 minutes, they should be able to use AI.
Larpy

PR #2281 is a strong base: one hard decision (what is safe to retry without paying twice), real tests already in the repo, and small enough to cut to 90 minutes. Here is the plan.

Assessment planDraft
Task
Retry failed payouts safely
Level
Mid-level backend
Time
90 minutes
They build
The retry logic in PayoutClient.send, with the queue and storage stubbed.
Scored on
Correctness, safety under retries, rate limits, tests they add, how they direct AI.
Answer key
Acme’s merged code from #2281.
Candidate repo payouts/ errors.py processor_fake.py added tests/ pyproject.toml and 7 more Only you see hidden_tests/ 9 answer_key/ #2281

Use all the AI you want. We’ll show you how.

booking.ts
Written by them 9 linesWith Claude Code 5 linesExplained on the call 3 of 3

It records the task folder. Nothing else.

  • No keystrokes and no other apps.
  • One file candidates can read before they run it.
  • A gap shows as a gap. Nothing is filled in.
  • Hidden tests run after they submit.
On the record
Written with Claude Code14 lines
Explained on the code call3 of 3
Q2“If confirm throws, finally still releases it.”

Questions

Still have a question? Book a demo

How do we know the work is really theirs?

The code call checks they understand what they handed in. During the take-home, Larpy records the task folder and nothing else, and candidates can read exactly what it records before they start.

Can candidates use AI?

Yes. Larpy records which prompts wrote which code, then asks them about it. What they cannot explain counts.

Which languages work?

TypeScript, JavaScript, Python, Go, Java, Rust and Ruby.

Do candidates need to install anything?

One command, or a workspace in the browser.