Decide what you want to learn
Pick one question before you pick a failure: can the on-call engineer find the right dashboard, does the runbook still match the system, do two teams hand over cleanly. A gameday with no question ends with a list of everything.
Keep it away from production
Give each participant or pair their own account with the failure already in place. Nobody waits on a shared environment, and nobody's mistake spills into a colleague's exercise or a customer's traffic.
Run it against a clock
Announce the time limit and keep to it. Pressure is part of what you are rehearsing, and a fixed end makes the debrief happen while the details are fresh.
The debrief is the deliverable
Close with a short, blameless review.
- What did people check first, and was it the right first check?
- Where did the documentation help, and where did it send someone the wrong way?
- Which one change to tooling, alerting or runbooks comes out of today, and who owns it?