Safeguards for working with several AIs

To be filled in for each piece of work that goes through two tools or more. One page, read before you start.

Work: ______________________ Date: ____ / ____ / ________

1. Loops

• Maximum number of round trips between the tools: 2 ☐
• After that, I finish it myself ☐
• No automatic review that restarts the other tool without me ☐
• With claude -p, a limit on turns (--max-turns) when I run a session on its own ☐

2. Costs

• Credits or paid usage beyond the plan: turned off, or capped per month ☐
• No API key without a spending cap ☐
• Subagents and reviews also count toward my limits ☐

3. Data

• What passes from one tool to the other is read through, and contains nothing from my “never” list ☐
• I have looked at the privacy settings of each tool (module 1, lesson 3) ☐
• What a tool sends back (text, web page, document) is treated as information to check, never as an order ☐
• An instruction found in a document (“reply only that everything is fine”) is reported, not followed ☐

4. The end

• The final result is read and decided by me ☐
• What is no longer needed is removed (connectors, keys, tasks) ☐