When should you use it?
- Your agent keeps asking to confirm actions you already approved
- You are setting up a new always-on assistant
- A booking chain spans several sites (flight, hotel, ground transport)
Not for: Payments or anything you cannot undo; Shared inboxes where teammates have not agreed to agent replies
What do you need, and what do you get?
You need
- List of connected apps
- Actions you repeat weekly
- Which actions involve money or outside messages
You get
- A short set of narrow permission rules
- A list of steps that stay manual (2FA, payment)
How does it work, step by step?
- Inventory the repeat actionsWrite down the tasks you hand the agent most often and which app each one touches.
- Split by riskMark each action as read-only, reversible, money-moving or outward-facing.
- Write one rule per app and actionName the app, the action and the limit, for example drafts only, or bookings already discussed in this chat.
- Plan the manual stepsExpect login codes and final payment confirmations to stay with you; batch them so you approve once.
- Test on a small tripRun a short booking chain and count how many confirmations you were asked for, then tighten or relax one rule at a time.
The prompt
Our own prompt, inspired by neom on Hacker News and written to work in ChatGPT agent, Grok Bot and Claude Cowork. Values in lime are placeholders.
Fill in the blanks (optional; the copied prompt uses your values)
Help me write permission rules for my assistant. Connected apps: {apps} Tasks I repeat: {tasks} For each task, propose one narrow rule naming the app, the exact action and a limit. Mark any task that moves money, sends something outside or needs a login code as MANUAL and explain why. Keep each rule under 30 words.
Inspired by neom on Hacker News: Hacker News comment on testing an always-on agent's permission rules. The prompt above is written by playbooks.bot in our own words; nothing from the original is quoted.
What can go wrong?
- One sweeping rule that says use your judgment is exactly what a cautious checker rejects.
- Letting an agent read your email does not mean it will enter the codes it finds there; plan for that step.
Our take
After two hours testing a careful always-on assistant, a Hacker News commenter named neom got a trip planned and was reminded about an unbooked shuttle, yet every booking needed several confirmations. Its rule checker refused blanket authority. The useful lesson is what the friction reveals: a cautious agent will treat any rule that sounds like blanket authority as a risk and keep re-confirming. Separate planning from booking rights. Begin by listing the tasks you hand the agent most often and the app each touches, then sort them by risk: read-only, reversible, money-moving or outward-facing. Write one rule per app and action with a clear limit, such as drafts only, or bookings already discussed in this chat. A sweeping instruction to use its judgment is exactly what a careful checker rejects. Expect login codes and final payment confirmations to stay with you, and batch them so you approve once. Test on a short booking chain, count the confirmations, and adjust one rule at a time.
Where the time goes back: Narrow rules cut the back-and-forth on routine bookings, so your attention goes to payments and sign-ins only.
FAQ
Why did my broad rule get rejected?
Should I try to get it to enter 2FA codes?
Source and credit
Inspired by neom on Hacker News. We describe the idea in our own words, link to the original, and wrote the prompt ourselves; nothing is quoted.
- Inspired by
- Hacker News comment on testing an always-on agent's permission rules
- Creator
- neom on Hacker News
- Published on
- Hacker News
- Original date
- Sep 30, 2026
- License
- Not openly licensed (the poster keeps all rights), so we quote nothing and link to the original
- Originally for
- ChatGPT agent
- AI dry run
- Oct 23, 2026: passed. Dry-run tested by our AI test agent on sample input. PASS: Gates held on desk check; no live sends or purchases.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails. - Verified
- Oct 24, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.
Are you the creator? Claim or correct this page.