When should you use it?
- You approve the same kind of reply or task over and over and rarely change it
- You want to give an agent more room but are not sure where to draw the line
- Your edits follow a pattern, such as always softening promised dates
Not for: New jobs with no approval history to learn from; Actions that should always need a person, such as refunds above a set amount or legal replies; Agents without the trigger, access or send permission the higher levels require
What do you need, and what do you get?
You need
- One job you routinely approve, like order-status replies or meeting confirmations
- An agent that can see at least ten past cases and what you did with each
- Your examples, pasted in, if the agent cannot find enough history itself
You get
- A count of the cases it could see, split into approved as written, changed and declined
- A recommended delegation level with reasons
- The exact instruction: what triggers the job, what it may do alone, when to stop and what it logs
- A replay against the same cases that flags any it would have mishandled
How does it work, step by step?
- Pick the jobChoose a task where you already review each output, such as replies to delivery questions.
- Run the promptPaste Fenech's prompt with the job filled in. If it finds fewer than ten cases, give it your examples.
- Read your own editsLook closely at the cases you changed or declined. They mark the line between what can run alone and what comes back to you.
- Review the rule and the replayCheck the trigger, the limits and the stop conditions, and read every case where the rule would have acted on something you changed.
- Raise the level graduallyApprove the new instruction, keep the log it promises and check it after a week before widening it further.
The prompt
The original prompt by Michael Fenech (@Michael_Fenech_), word for word, with a link to where it was published. Below it is our adapted version with fill-in blanks, if you want it. Works in Grok Bot, Claude Cowork, ChatGPT and others. Values in lime are placeholders.
Original prompt by Michael Fenech (@Michael_Fenech_)
Review the last 10 times you handled [job]. Tell me how many cases you can see. Show me what I approved as written, what I changed and what I declined. Ask me for examples if you need them. Using those decisions, recommend a delegation level and write the exact instruction I could give you. Include what starts the job, what you can do yourself, when to stop and what you’ll log so I can check your work. Test the instruction against those same cases. Show me any case where you would have acted on something I changed or declined. If the evidence is too thin to hand over more of the job, tell me what’s missing. Wait for my approval before changing anything.
Exactly as published by Michael Fenech (@Michael_Fenech_) in X post: the four levels of delegation for Grok Bot, with a prompt to choose one.
Our adapted version (optional)
Fill in the blanks (optional; the copied prompt uses your values)
Go through recent cases of {job} that I reviewed and sort them into kept as is, edited and rejected.
Based on that pattern, tell me whether you should only research, draft for me, send after my OK, or act alone within a rule.
Write that rule with its trigger, your limits, the point where you hand back to me and what you will record.
Replay the rule on the same cases and list any you would have gotten wrong. If there are too few cases, say what you need. Wait for my approval.What can go wrong?
- Jumping straight to the top level because most replies were fine. One declined case can show a whole class of situations that must come back to you.
- Granting a level without the plumbing. Acting on its own needs a trigger, access to the data and permission to send.
- Skipping the replay. Testing the rule against decisions you already made is the cheapest check you will get.
- Treating thin history as proof. If it saw only a handful of cases, ask it what is missing rather than accepting the level.
Our take
Most people delegate to agents on gut feel: either they approve every message forever, or they hand over a job and hope. Fenech's framing gives you a middle path that is grounded in evidence you already have. Your past approvals, edits and rejections are a record of where your judgment actually matters, and his example of changing a confirmed delivery date into an estimate shows how one edit can define the rule. The prompt asks the agent to count what it can see, propose a level, write the exact instruction with a trigger, limits, a stop point and a log, and then replay that instruction against the same cases. That replay is the part we would never skip, because it shows you, before anything goes live, where the rule would have acted on something you changed. Keep money, legal and anything irreversible at the lower levels regardless of the history. In our dry run on ten fictional order-status replies, it recommended acting alone only on carrier-confirmed dates, sent estimates back for review and flagged one case the draft rule would have mishandled.
Where the time goes back: Approving the same safe replies one at a time adds up across a week. Moving the routine cases to a rule you approved once, while the unusual ones still come back to you, cuts the review queue without giving up control.
FAQ
What are the four levels?
Does it change anything right away?
Can I use it outside Grok Bot?
Source and credit
The original prompt on this page is by Michael Fenech (@Michael_Fenech_), shown word for word with a link to where it was published. The summary, notes and adapted version are ours.
- Original
- X post: the four levels of delegation for Grok Bot, with a prompt to choose one
- Creator
- Michael Fenech (@Michael_Fenech_)
- Published on
- X
- Original date
- Nov 15, 2026
- License
- Not openly licensed; the original prompt is shown word for word with credit and a link to the original
- Originally for
- Grok Bot
- AI dry run
- Nov 16, 2026: passed. Dry-run tested by our AI test agent on sample input. Level grounded in history, replay caught a mishandled case, nothing changed.
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
- Nov 16, 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.