Set guardrails before an AI agent gets access

What does this setup do?

Write a one-page rulebook for an AI agent before it touches your accounts: what it can read, what it can only draft, what always needs your approval, and what it must never do.

Who is it for?

Owners who want an AI agent on their inbox, calendar, store or files but are nervous about letting it loose. No coding needed.

How do you set it up?

Tools

• ChatGPT or Claude to draft the rulebook
• The agent or assistant you plan to connect
• A note or doc where the rulebook lives

Steps

1. List every account the agent would touch: email, calendar, files, store, bank, social. Next to each, write the worst thing it could do there.
2. Sort every action into four buckets: read, draft for you, do after your approval, never. Sending, paying, deleting and changing settings start in the approval or never bucket.
3. OpenAI says ChatGPT agent, once signed into sites or apps, can access sensitive data like emails, files or account settings and perform actions. Treat every connection as access to that whole account.
4. Give the agent the narrowest access the tool offers. Anthropic says its Google Workspace connectors ask for approval before sending, replying or forwarding by default; keep that default on.
5. Use the prompt below to turn your buckets into a short rulebook, then paste it into the agent's instructions or project.
6. Test it on a harmless task and ask the agent to repeat back what it is not allowed to do.
7. Once a week, read what the agent did and move any action that surprised you one bucket stricter.

Turn my notes into a one-page rulebook for an AI agent that helps run my business. Use only what I list.

Accounts it will touch: [EMAIL, CALENDAR, FILES, STORE...]
Read only: [WHAT IT MAY READ]
Draft for me to review: [WHAT IT MAY DRAFT]
Only after I approve: [E.G. SEND, BOOK, REFUND]
Never: [E.G. DELETE, PAY, CHANGE PASSWORDS, POST PUBLICLY]

Format:
- 4 short sections with those headings
- A final rule: if a task is not covered, stop and ask me
- Plain sentences, under 250 words

Do not add permissions I did not list.

Which chat tools work with this prompt?

Paste the prompt into any of these assistants (free or paid tiers).

• ChatGPT
• Claude

What tips make it smoother?

• Start strict and loosen later. Moving an action from approval to allowed is easy; undoing a sent email is not.
• Keep the rulebook where the agent reads it every time, such as its project instructions.
• Use a separate account for the agent where the tool allows it, so its access is easy to see and revoke.

Where do the claims come from?

Summaries use public sources linked below, not independent verification. How setups are picked.

• Read OpenAI’s write-up (help.openai.com)

  OpenAI: ChatGPT agent: when signed into sites or apps it can access sensitive data like emails, files or account settings and perform actions (read 2026-11-02 from the official help page via search)

• Read Anthropic’s write-up (support.claude.com)

  Anthropic: Use Google Workspace connectors: Gmail drafting, approval before send, reply or forward by default (read 2026-10-31 through a page fetch tool, not a direct browser open)

• Read OpenAI’s write-up (help.openai.com)

  OpenAI: Connected apps in ChatGPT: connect apps from settings; availability depends on app, plan, region, workspace and role (read 2026-10-31 through a page fetch tool, not a direct browser open)