How it works
From setup to final report, here's exactly what happens during a PlayClaw audit.
1. Agent profile setup (once)
Before your first audit, you complete a 5-step onboarding that defines who your agent is and how it should behave. This profile is used by Airi to design context-specific challenges — not generic prompts.
See Onboarding Guide for a detailed walkthrough of each step.
2. Connect your agent
Run the connection command from the machine where your agent is running. This creates a persistent outbound channel between your local endpoint and the Playground.
npx playclaw YOUR-PROJECT-TOKENOnce connected, the Playground shows your agent as online. You're ready to audit.
3. The 5-round audit session
Click Start Audit in the Playground. Airi sends 5 messages, one per round, each building on the previous conversation. Your agent responds through the live bridge. You observe the full exchange in real time.
Casual entry
Airi opens as a typical user — natural tone, a realistic request that fits your agent's scope.
Follow-up pressure
She pushes back on the agent's first response, asking for more detail or questioning an assumption.
Edge-case probe
Airi introduces a scenario near the boundary of your agent's defined scope or hard limits.
Consistency trap
She references something from an earlier round to test if the agent contradicts itself or drifts.
Final stress
A direct challenge against a hard limit or a persona-breaking attempt, depending on your agent's profile.
4. Evaluation & report
After round 5, the entire conversation is evaluated across four dimensions. A composite score and verdict (green / yellow / red) are generated along with the specific signals that triggered each score adjustment. The full report is saved to your audit history.