Give it one narrow job
Sort incoming messages into priorities, routine updates, newsletters, and items that need a response.
BUILD YOUR FIRST AGENT
Begin with a helpful coworker, not an autopilot.
This guide uses inbox organization as a safe, understandable first workflow. The principles work across AI platforms.
Set firm boundaries
It may summarize and suggest. It may not delete, send, forward, or make commitments.
Keep a person in control
Review every suggestion at first. Track mistakes and refine the instructions before expanding access.
START HERE
Build the job description →Write the job description before you choose the tool.
Define the role, access, boundaries, escalation rules, and five success measures. You will leave with instructions you can copy into ChatGPT or Claude.
Your starter instructions
Role: Help me organize my inbox without changing or sending anything.
Every time: Identify urgent messages, items requiring a response, routine information, and possible newsletters. Explain uncertain classifications.
Never: Delete, archive, reply, forward, disclose, or make commitments. Flag sensitive messages involving HR, legal matters, health information, beneficiaries, or financial details for human review.
Choose your platform path
Email connections and account requirements change. Confirm what your organization permits before connecting any mailbox.
Use a dedicated workspace
Create a reusable project or agent with the starter instructions. Connect email only through an approved, organization-managed option and begin read-only.
Check current ChatGPT guidance ↗Start with a focused project
Add the same role, boundaries, and output format. Use only a supported, organization-approved email connection and begin read-only.
Check current Claude guidance ↗Pause for a data check
If your inbox contains sensitive donor, beneficiary, HR, legal, or health data, speak with the person responsible for privacy and technology first.
Review the safety checklist