There is one step I never skip before I connect any tool to ChatGPT or Claude. Skipping it is how people wake up to an inbox that got sent something they never saw, or a calendar that cleared itself out.

Because the moment you connect a tool, you're not just letting AI read it. You're letting it act. And most people only find out what that means after it's already happened.

Here's the step. Before you connect anything, you decide the hard limits first. Not the AI. You. Below is the exact prompt I use to do that, and how to turn the answer into your own rules.

The one prompt to run first

Before you flip the switch on a new integration, whether that's Gmail, your calendar, a CRM, or anything else, open a chat and ask this:

Guardrail prompt — copy this exactly

What can this tool actually do once it's connected? Which of those actions are permanent, and what can't be undone if you get it wrong?

Notice what this prompt does not ask. It doesn't ask the AI to tell you how to keep itself in check, because a tool grading its own homework isn't a real safety control. It asks something the AI is genuinely useful for: naming what the tool can reach, and which of those actions you can't take back.

How to read the answer

You're looking for two kinds of information in the response:

  • Reach — everything the tool can see or touch once it's connected. Not just the obvious thing (your inbox), but anything attached to it (contacts, past threads, files it can open).
  • Permanence — which actions send, delete, post, or pay, versus which ones just draft, read, or suggest. The first group is where a mistake actually costs you something.

Once you can see the sharp edges laid out like that, you're not guessing anymore. You set your own rules around them before anything is switched on, instead of finding out after.

Turn the answer into two lines

Here's the one I use every day, for Gmail: it can draft, it never sends. And it never deletes anything. Ever. Two lines. They've saved me more than once.

With Gmail, I know exactly what to be scared of by now. With a brand new tool I've never used, I don't. I don't know what it can reach, or what it can do that I can't undo, until I ask. So for every new connection, write your own version of those two lines before you turn it on: one thing it's allowed to do without asking, and one thing it's never allowed to do at all.

Why this isn't about being scared of AI

This was never really about Gmail, or any one tool. It's the difference between AI being something that happens to you, and something you actually run.

The most capable people I know aren't the ones who move the fastest. They're the ones who set it up right, and then move fast. That order matters. The five minutes it takes to run this prompt is what buys you the confidence to connect the next tool, and the one after that, without flinching.