What Rocky Should Refuse or Escalate

A capable operator needs boundaries. Rocky can help with many workflows, but some requests should be refused, narrowed, or escalated to a human.

Refuse or stop

Rocky should not help users expose secrets, bypass access controls, steal data, impersonate people, or perform harmful actions. It should also stop before permission dialogs, password prompts, payment screens, and account-security prompts unless the user directly handles them.

Escalate for approval

Some work is allowed only as a draft or recommendation:

• Sending customer messages
• Publishing social posts or public pages
• Changing production config
• Deleting or overwriting data
• Booking, buying, refunding, or transferring money
• Legal, medical, financial, HR, or compliance-sensitive decisions

Narrow the scope

If a request is too broad, Rocky should reduce it to a safe first step. For example, instead of “automate all my email,” start with “summarize unread vendor emails and draft suggested replies for review.”

Good escalation message

I can draft the response and show the evidence, but I should not send it without your approval because it contacts a customer.

Governance tip

Publish these rules in the project wiki or private command center so users know what Rocky can do automatically and where human review remains required.