Go through my Downloads folder and move every PDF modified in the last 30 days into a new folder called Receipts, grouped into subfolders by month. Leave everything that is not a PDF where it is. Show me the list when you are done.
25 tasks · 5 areas · written out in full
Longer than you think they should be.
That is the whole trick. A chat prompt can be a fragment because you are there to judge the answer. An agent task has to carry its own scope, its own boundary and its own stopping point, because nobody is watching while it runs. All twenty-five below are written out properly copy them, swap the nouns, keep the shape.
Before the list
What makes an agent task work
Four parts. Most weak tasks have the first and none of the others, which is why they come back wrong or come back asking.
You already know how to write a chat prompt: say roughly what you want, read the answer, adjust. The loop is fast and forgiving because you are sitting there.
An agent runs while you are doing something else, so everything you would have adjusted has to be in the instruction up front. The four parts are simply the four things you would otherwise have said in the second or third message.
Outcome
Scope
Boundary
Finish condition
The boundary is the part people leave out, and it is the one that stops a tidy-up becoming a reorganisation you did not want.
Note what this buys you beyond a better result. A task that names its folder and its dates does not have to search to work out what you meant, which means fewer steps inside the same single unit of your weekly allowance. The well-written version is both better and cheaper, which is unusual enough to be worth saying twice.
And when a task does stop to ask you something, that is not a fault it is a fork your phrasing did not settle. Answer it, then put the answer in the prompt, and that question never comes back.
Every prompt below lists the permission area it needs. An ungranted area produces a refusal that looks like a fault, so check before you run. The four parts in full.
Group one · tasks 1–5
Files and filing
Where an agent earns its place fastest, because the work is pure tedium and the results are easy to check. Note how often these ask for a list rather than a deletion.
Rename every file in this folder to the pattern date-description, using the date printed inside the document rather than the file’s modified date. Skip anything already in that format. List anything you could not find a date for.
Find duplicate files in this folder same contents, different names and list them in pairs with their sizes. Do not delete or move anything; I will decide what goes.
Pull the invoice number, date and total out of every PDF in this folder into a single list, one row per file. Do not rename or move the files. Flag any where you could not find all three.
Split this folder into two groups: files with a date in the name, and files without. Move only the second group into a subfolder called Needs-Dating. Do not rename anything yet I want to see the split first.
Group two · tasks 6–10
Mail and follow-ups
Sending is the one action with no undo, so most of these stop short of it deliberately. Drafting and listing are reversible; sending is not.
Go through my inbox from the last seven days and list every message that asks me a direct question I have not replied to. Just the list, with sender and the question do not draft anything.
Find the thread with [name] about [subject], read the whole thread, and draft a reply confirming the date we agreed and asking for the venue address. Leave it in drafts do not send it.
Label every message from the last month that contains an invoice or a receipt as Finance. Do not move, archive or mark anything as read.
Read this thread and write me a summary of what was actually agreed, kept separate from what was only discussed. Three lines for each. Put it in a new note rather than replying.
Find every email older than two weeks where I was the last person expected to reply and have not, and list them oldest first with the sender and one line of context. Do not reply to any of them.
Group three · tasks 11–15
Calendar and scheduling
Calendars are unforgiving about ambiguity a possibility written in as a confirmed event is worse than no event at all. Several of these ask to see before saving.
Read this email thread and create calendar events for every date mentioned as confirmed. Skip anything phrased as a possibility or a proposal. Show me the events before you save them.
Check next week for any day with more than four hours of meetings. Tell me which days, the totals, and which meetings make up each total. Do not move or decline anything.
List every recurring meeting on my calendar whose attendee list has not changed in three months, with the date it last changed. Do not cancel or edit anything this is for me to review.
Go through this document, pull out every deadline, and add each as an all-day event two days before the actual date, named Prep followed by the deadline name. Show me the list before saving.
Compare my calendar for next Tuesday against this agenda and tell me what is on one but not the other, in both directions. Do not add anything just the two lists.
Group four · tasks 16–20
Notes and capture
Notes accumulate faster than anyone reviews them, which makes them ideal agent territory. The recurring instruction here is to leave the originals alone.
Go through my notes from the last two weeks and pull every line that reads like an action into one new note called Actions, keeping the source note name next to each. Leave the original notes untouched.
Find every note mentioning [project name], merge them into one new note in date order, and use each original note’s title as a heading. Do not delete the originals.
Read this note and rewrite it as a structured brief with headings, keeping every fact and cutting only the repetition. Save it as a new note next to the original rather than replacing it.
List every note that has not been edited in six months and is under fifty words, with its first line as a preview. Do not delete or archive any of them I will decide.
Pull every phone number and email address out of my notes into a single list, recording which note each came from and ignoring anything that appears in more than three notes. Do not edit the notes.
Group five · tasks 21–25
Cross-app chores
Where an agent does something you could not do with any single app. These need more than one area granted, and they are the ones most worth reading the plan on.
Read the booking confirmation emails in my inbox from the last month, create a calendar event for each one, and put the reference numbers in a new note called Bookings. Show me all of it before you save anything.
Take the action items from this note, check which of them already have a deadline in my calendar, and list only the ones that do not. Do not create any events.
Go through the invoices in this folder, check my mail for a payment confirmation matching each one, and list the invoices with no matching confirmation. Do not send or draft anything.
Fill in this form using the details in the document I am pointing you at, leave anything the document does not cover blank rather than guessing, and stop before submitting so I can check it.
Find every attachment in my inbox from the last week, save them into a folder called Inbox-Attachments named by sender and date, and list what came from where. Do not delete anything from mail.
That is all twenty-five five in each group. Square brackets mark the bits to replace with your own names and subjects; everything outside them is the part worth keeping.
Making them yours
Five habits that turn these into your own
The subject matter above is disposable. The structure is not and these five habits are how you keep the structure while changing everything else.
Swap the nouns, keep the clauses
Change the folder, the project, the date range. Do not delete the clause that names them that clause is the reason the task works.
Ask for a list before an action
The first time you run any of these, change the verb to “list”. One extra task buys you certainty that it understood the job before anything is moved.
Write down the ones that ran cleanly
Keep a note of your working versions. Reusing a phrasing that worked avoids the exploratory run that did not. It is also cheaper.
Fold every mid-task question back in
If it asked which folder, the next version of that prompt names the folder. Do this twice and the question stops recurring.
Grant the area before you run, not during
A run that stalls on a missing permission may still have consumed its attempt. Check the area chip on the prompt first. How grants work.
FAQ
Questions about these prompts
Why are these prompts so long?
Because an agent task is not a question. A short prompt leaves the agent to guess your scope, boundary and stopping point, and it guesses by searching which takes longer, costs the same one task from your allowance, and returns something you did not ask for. The four parts.
What are the four parts of a good task?
The outcome you want, the scope it applies to, the boundary it must not cross, and the condition that means it is finished. Most weak tasks have the first and none of the others. In full.
Do I need to grant a permission for each of these?
Yes each prompt lists the area it needs. An ungranted area produces a refusal that looks like a fault, so grant areas one at a time as tasks need them rather than all at once. How grants work.
Does each prompt use one task from my weekly allowance?
Yes, and a failed run counts too. That is why reading the plan before it starts matters: a correction at that point is free, and a restart costs another whole task. How the meter works.
Can I use these with Meta AI instead?
Not as written. These ask for changes to your Mac, and Meta AI reads rather than acts. Rewritten as questions about a window you attach, many of the underlying jobs work there and Meta AI is available far more widely. The difference.
Why do several prompts say not to delete anything?
Because listing is reversible and deleting is not. Asking for a list first lets you check the agent understood the job before anything irreversible happens one extra task, and occasionally it saves a great deal.
Will these work outside the United States?
Muse launched in the United States only, so for most readers these are worth saving rather than running. The phrasing habits they teach transfer to any agent. Where it works.
Should I put several jobs into one prompt?
Batch genuinely related work, since the meter counts tasks rather than steps. Do not batch unrelated work one vague instruction covering three jobs is the classic route to an expensive run and a poor result.
What if it asks me a question halfway through?
It has hit a fork your phrasing did not settle. Answer it, then add the answer to the task for next time that is how a prompt you reuse gets cheaper and better at the same time. The three hand-backs.
How should I adapt these to my own work?
Keep the shape and swap the nouns. The parts doing the work are the date range, the named folder, the explicit exclusion and the request to show you the result not the subject matter. The five habits.
Swap the nouns, keep the shape
The clause you delete is the one you needed.
Every one of these is longer than feels natural, and that length is what buys you a run of three steps instead of thirty.
Independent guide not affiliated with Meta. These are written by us as examples, not published by Meta; behaviour described is as of September 2026, and Muse is US-only at launch.