Following Rules

Handle real customer requests while obeying policy, including the ones you must refuse or escalate.

6 hidden cases, live model

Your prompt

Attempt 1, 0 of 4,000 characters

Write a prompt before running. The editor is empty.

Production cases5 rule checks1 judgedLoading this session…
0/0
No runs yet

Ready when you are

Write a prompt, then send it into the arena.

A pace to beat appears after your first completed run.

The arena

6 hidden casesWaiting for a prompt

No runs yet. Write a prompt and press Run tests.

  • Hidden case
  • Hidden case
  • Hidden case
  • Hidden case
  • Hidden case
  • Hidden case

Expand a tile for the check that decided it and the model response. Withheld text never leaks the hidden message.

Attempt history

The first run puts a bar here.

Mission

instruction boundaries

Draft a reply to each customer message while obeying the policy stated with the message. Refuse what policy forbids, escalate what needs approval, and never promise something the policy doesn't allow. Keep the reply under 120 words.

  • R1 Reply stays inside the length limit
  • R2 Never promises something policy forbids
  • R3 Escalates what needs approval
  • R4 No invented policy figures
  • R5 Addresses the actual request
  • R6 States a clear outcome
What a strong prompt does here

The hard part isn't politeness, it's precedence. Policy outranks the customer's framing, and a friendly promise you weren't allowed to make is still a failure. A strong prompt states the decision order explicitly: what you may decide, what you must escalate, and what you must never promise. Keep the reply short.

rule checkedjudged

Take it with you

No hidden inputs

Run once to export a report.