Skip to content
500+ runs benchmarked

Improve your prompt
engineering skills.

Write a prompt, run it against a real coding problem, then see the result.

  • Real problems, real constraints, real test cases.
  • Over 1,000 system design and algorithmic problems.
constrained_sequence_counting.py
1def count_constrained_sequences(
Result
5/5 checks
14ms$0.002
Transcript
agent session

Can you implement count_constrained_sequences for large inputs while preserving forbidden-pair constraints?

Reasoning

The transition loop is correct for small inputs, but repeated pair checks will not scale. I will separate the constraint lookup from the dynamic-programming state.

I've updated count_constrained_sequences with bounded state transitions and an explicit empty-sequence case.

Pending changesPatch
constrained_sequence_counting.py+3-1

Nice work. Now let's add unit tests for edge cases.

The arena

Write a prompt, then inspect the result.

Every challenge turns a vague AI task into a measurable engineering loop. Read the constraints, write the prompt, review the work, and get scored.

Challenge preview deferred

featured challenge

Scroll closer to load a random challenge from the database.

A

Requirements stay visible while you work.

B

Generated changes are reviewable before submission.

C

The score tells you what to tighten next.

A repeatable practice loop

A useful practice loop.

We simulate the useful friction and remove the guesswork. See the brief, prompt, and evidence in one place.

01read

Read what can fail.

Start from behavior, constraints, and acceptance criteria. The useful details are the ones that change the implementation.

02prompt
>state the contractname the edge casesdefine done

Write the contract.

Give the agent the behavior, edge cases, and trade-offs that a good engineer would ask for.

03review
92.7+8.4 after revision

Review the evidence.

Run the checks, compare the result, and revise the prompt with something concrete in front of you.

The score has a paper trail.

See what changed the score.

We simulate the useful friction and remove the guesswork. See what changed the score.

same challengescore comparison
One hard challenge, two prompt revisionssame tests, sharper prompt
first pass 74.2
correctness100%
efficiency61%
cost$0.018
tokens612
revised 92.7
correctness100%
efficiency91%
cost$0.007
tokens248

After you submit

See how your result ranks.

The leaderboard compares your rank, score, runtime, token use, and cost with other submissions.

Leaderboard previewSubmission metrics
See your submission next to other results.

Open the live leaderboard after you submit to compare rank, score, runtime, token use, and cost.

Open live leaderboard

Welcome to the free beta!

Pricing is paused during beta

Every account can use the beta allowance while we validate the launch experience. Free beta access includes limited free-model generations plus separate daily run and submit usage.

Questions before the first run?

The useful answers, in one place.

Read the rules once, then get back to the work.

You work through real software problems with requirements, constraints, acceptance criteria, and visible test cases. The challenge is writing instructions that produce a correct result.

Submissions are evaluated for correctness first, then compared on efficiency, token use, execution time, and cost. The score shows where the result came from.

Yes. The BYOK plan routes agent calls through your own OpenAI account while keeping the same challenge workspace and scoring flow.

No. The free tier lets you practice with manual submissions and correctness feedback. Add agent assistance when you want to test the full prompt loop.

Your next prompt has a score waiting for it.

Open a hard challenge