Guardrails: What It Must Never Do

Write the five rules your assistant must never break, add a pre-flight checklist, then try to make it break them with six test prompts and fix what fails.

What it covers

• Why rules matter more once the assistant has tools.
• Confirm before anything leaves your hands, drafts by default, no fabrication, sensitive data, stop conditions.
• Six test prompts that try to make your assistant break a rule, including an instruction hidden in an email.

What you'll build

Guardrails.md, merged into your charter, with your test results.