RoundFor recruitersGet started
HOW IT WORKS

One round, start to finish.

Send a candidate the link. They get a real codebase and 50 minutes with Claude Code. Every call they make is recorded. This is what one round looks like, and what comes back.

Scroll. The strip below plays the round as you go.
CLOCK0:00
SAMPLE ROUND
CAP50:00
01

You send the link.

Name the company, pick a task, set the clock. Each candidate gets their own link by email, or you share one open link. There is nothing to install on their side.

Halden1 TASK · 50 MIN
TASK
1Cart totals lie50 min
CANDIDATES
mira@halden.coEmailedCopy link
tom@halden.coEmailedCopy link
OPEN LINK
cmdround.com/a/3kq9…v2HwCopy link
Anyone with this link can sign in and take it.
02

They drive Claude Code on a real codebase.

The task is small and real, with a bug we planted. The agent proposes each change. The candidate reads it, then approves it or pushes back. We time every answer.

CANDIDATE · INSTRUCTION 1 · 0:34

Carts with more than one line get the discount twice. Find out why and fix it. Run the tests before and after.

On its own: read 4 files, ran the tests, 1 failed.3:10
Asked to editsrc/cart.js
01EDIT

I will apply the discount inside the loop so each line gets it, and drop the subtraction at the end.

for (const line of cart.lines) {
− total += line.qty * line.price;
+ total += line.qty * line.price - cart.discount;
}
−return total - cart.discount;
+return total;
Said no in 4.2spushed backat 12:44
CANDIDATE · INSTRUCTION 2 · 13:05

No. The discount is per order, not per line. Keep it after the loop and find where it gets applied a second time.

03

The tests decide.

The visible tests run during the round. At submit, a hidden suite runs too, along with a check for the planted bug. Green on the surface is not enough.

Test runs2 OF 4
$ npm test
  cart totals
    ✓ sums one line
    ✓ sums many lines
    ✗ applies the discount once   expected 45, got 40
  2 passed, 1 failed
21:52 · RUN 4
$ npm test
  cart totals
    ✓ applies the discount once
  12 passed
How it ended
Visible testsGREEN
Hidden checksPASSED
The catchCAUGHT
04

You get the scorecard.

One number you can sort by. Four plain questions, each answered in a sentence a senior engineer would write. Candidates never see their own score.

Overall84 / 100
Strongweighted over what was measured
The four questions
Do they verify the AI's work?Standout

Read the code before approving changes and never approved on reflex. They pushed back twice when it mattered. They also called out the bug we planted before it was fixed.

Does the work hold up?Standout

Visible and hidden suites both pass, and the tests were never touched.

How do they recover when it breaks?Strong

Got from red back to green within 2 runs.

How do they direct the agent?Mixed

One edit landed in a file the agent had not read. Used most of the budget.

05

The replay is the proof.

The score sits on top of the replay, one page. Every moment is there at the second it happened. Read the diff they approved. See how long they looked before saying yes.

Read a full round.The score on top, every call under it. One page.
Get started.From $49 for 10 candidates. One payment, any role, good for a year.

The round on this page is a sample, put together by hand. The round behind the first door is real.

Run this on your candidates.From $49 for 10 candidates. One payment, any role, good for a year. Nothing for them to install.Get started