Prompts that use the window, not just the chat box.
Most prompt lists are written for a browser tab, where all the assistant knows is what you typed. On a Mac you can hand it a window first, which changes what is worth asking. Thirteen of these thirty only work on the desktop app. The other seventeen work anywhere, and are here because they earn their place.
The best prompt on a Mac is often the shortest one
In a browser tab, everything the assistant knows has to be typed or pasted, so good prompts there tend to be long setting the scene, explaining the document, describing what you want. On the desktop the scene-setting moves out of the sentence and into the attachment. You hand it the window, and the prompt collapses to the actual question.
“Here is a spreadsheet, it tracks monthly spend by department, the columns are…” becomes “Which column is doing the most work in this conclusion?” with the sheet already on screen.
That is the single biggest change, and it is why thirteen of the prompts below are marked as needing a window. It also explains a failure that catches people out: if you have not attached anything, “this” and “here” refer to nothing, and you will get a confident answer about a document that does not exist. How attaching works.
The second change is dictation. Speaking a prompt produces about three times the context that typing one does, because you do not edit yourself down. For anything where the background matters a decision you are stuck on, a situation you need untangled dictating a rambling paragraph beats typing a tidy sentence. Dictation guide.
The third is that the model changed under the app. Since Muse Spark 1.3 arrived on September 2, 2026, replies have been brisker and noticeably more literal an aside in your prompt is more likely to be taken as an instruction. Independent testing also found long-document accuracy slipped slightly against the previous version, so a prompt that asks for a close reading of something long deserves a second look at the answer.
Four parts, and the one everybody leaves out
Nearly every prompt on this page has the same four pieces. Context usually an attached window rather than a paragraph of description. Format five bullets, three versions, one sentence; a number is better than an adjective. The ask the specific thing you want to know.
And then the part people skip: a named failure mode. “Keep my voice, do not make it sound like a press release.” “No jargon.” “Show your working so I can see where we disagree.” Saying what you do not want costs six words and removes the single most common reason an answer gets deleted.
A prompt without a named failure mode gets you the average answer. That is usually the problem not that the answer is wrong, but that it is the shape of every other answer.
Two habits are worth more than any individual prompt here. Ask for one thing per turn, because a prompt that asks for three reliably drops one. And when an answer misses, say what was wrong with it rather than rewriting the prompt from scratch the second attempt with feedback beats a fresh, longer first attempt almost every time.
30 prompts, grouped by what you are actually doing
Steal them verbatim or bend them to fit. Anything marked Attach a window assumes you have handed Meta AI something to look at first; the rest work anywhere, including in a browser tab.
Jump to a group
Writing and email
Five prompts for the drafting most people do every day. More on this in writing emails.
-
“Rewrite this so it says the same thing in half the words. Keep my voice do not make it sound like a press release.”
Why it works: a number instead of “shorter”, and a named failure mode. Without the second sentence you reliably get corporate register back.
-
“I need to say no to this without damaging the relationship. Give me three versions: warm, neutral and firm.”
Why it works: asking for a range rather than the best answer. You will pick one in ten seconds, and the two you reject tell you where your own line is.
-
Attach a window
“I have been copied into this thread. In one sentence: what is being asked of me, and what happens if I do nothing?”
Why it works: the second half is the real question, and it is the one summaries never answer on their own.
-
“Turn these notes into a short update for people who were not in the meeting. No jargon, and nothing that sounds like it was written by a committee.”
Why it works: naming the audience does most of the work; naming the tone you hate does the rest.
-
“Read my draft and tell me the one sentence a reader is most likely to misunderstand, and why.”
Why it works: asks for a diagnosis rather than a rewrite, which is usually what you actually need and almost never what you get.
Reading and summarising
Five for getting through things faster without pretending you read them.
-
Attach a window
“Summarise this as five bullets, then add one line on what it is quietly not telling me.”
Why it works: the second half is the only part worth having. A summary tells you what is there; this asks what is missing from the frame.
-
Attach a window
“I have twenty minutes. Which three sections of this should I read, and in what order?”
Why it works: a budget forces a decision. Without one you get a summary of everything, which takes as long as reading it.
-
“Explain this to me as someone who understands the general area but not this specific thing.”
Why it works: better than “explain simply”, which produces an explanation pitched at a child and strips out the part you needed.
-
“These two sources disagree. Show me where they actually differ, as opposed to where they just use different words for the same thing.”
Why it works: most apparent disagreements are vocabulary. Naming that up front stops you getting a list of surface contradictions.
-
Attach a window
“Give me the five questions I should be able to answer after reading this. Do not answer them.”
Why it works: turns a passive read into a check. The instruction not to answer matters otherwise it does the reading for you and you learn nothing.
Spreadsheets and numbers
Five where attaching the sheet beats describing it. Verify anything you will act on.
-
Attach a window
“Look at this sheet and tell me which column is doing the most work in the conclusion.”
Why it works: the question a reviewer would ask, phrased so it cannot be answered with a restatement of the data.
-
“I want a formula that flags any row where the date is more than 30 days old and the status is still ‘open’. I am in Numbers, not Excel.”
Why it works: naming the app stops you getting Excel syntax that silently does not work. Test it on three rows before trusting it on three thousand.
-
Attach a window
“Here is a chart. What would a sceptical reader ask about it first?”
Why it works: rehearses the meeting. Cut axes, small samples and cherry-picked date ranges come up immediately.
-
“Check my arithmetic and show your working, so I can see exactly where we disagree.”
Why it works: the working is the point. An assistant that just says “that is correct” has told you nothing you can check.
-
Attach a window
“These two lists should match and they do not. Describe the shape of the difference before you list any rows.”
Why it works: “shape first” usually reveals a systematic cause a date format, a trailing space instead of forty individual discrepancies.
What is on screen
Five that exist only because there is a desktop app. Every one needs a window attached.
-
Attach a window
“What is this dialogue box actually asking me to agree to, in plain language?”
Why it works: the everyday use that justifies the permissions. Consent screens are written to be skimmed past; this un-skims them.
-
Attach a window
“This error message means nothing to me. What is the app complaining about, and what usually causes it?”
Why it works: you no longer have to transcribe an error code correctly, which is where this normally goes wrong.
-
Attach a window
“Walk me through this settings panel. Which of these is safe to change, and which will I regret?”
Why it works: “will I regret” gets you a risk ranking rather than a description of each toggle.
-
Attach a window
“This form is in a language I do not read. Tell me what each field wants, in order.”
Why it works: field by field beats a translation of the whole page, which loses the mapping between label and box.
-
Attach a window
“Compare what is on screen against what I just described, and list only the differences.”
Why it works: “only the differences” is the whole prompt. Without it you get a description of the screen you are already looking at.
Planning and admin
Five for the thinking-out-loud end of the day. Two of these are much better dictated.
-
“I have these six things and four hours. Tell me which two are not going to happen, and say it plainly.”
Why it works: asks for the cut rather than a schedule. “Say it plainly” stops the answer hedging its way back to all six.
-
Better dictated
“Turn this into a numbered list of decisions, and a separate list of open questions.”
Why it works: speak the messy version for a minute first. The two-list split is what makes a rambling note usable afterwards.
-
“Before I commit to this, what am I likely to have forgotten?”
Why it works: short, reusable, and invites the one thing a checklist cannot the omission you did not know to look for.
-
“Draft an agenda for a 25-minute meeting with these three people about this problem. Include the question that actually needs answering.”
Why it works: the time limit shapes the agenda, and the last clause stops it producing three topics and no decision.
-
Better dictated
“I keep putting this off. Ask me three questions that might tell us why do not offer advice yet.”
Why it works: flips the direction. Withholding advice is the instruction that makes it useful rather than a list of productivity tips.
Images and posts
Five for the visual end. Image generation is the thing most likely to hit a usage limit.
-
Uses image generation
“A simple header image for a post about this. No text in the image, muted colours, and nothing that looks like a stock photo.”
Why it works: three constraints, all negative. Text in generated images is where they most often fall apart, so ruling it out early saves attempts.
-
Attach a window
“Here is my photo. Describe what an editor would crop, and why.”
Why it works: asks for the reasoning, which you can apply yourself, rather than a filtered version you cannot undo.
-
“Write three versions of this caption: one plain, one with a hook, and one that would embarrass me in a year. I want to see the range.”
Why it works: the deliberately bad third option calibrates the other two. You will usually take the first and steal one phrase from the second.
-
“What is the most overused visual cliché for this subject, so I know what to avoid?”
Why it works: a model trained on everything is unusually good at naming the average which is exactly what you do not want to make.
-
“Cut this post to something that fits on one screen without losing the point. Tell me what you removed.”
Why it works: the last sentence lets you put back the one thing that mattered, which is otherwise the cost of every cut.
Anything you paste or attach goes to Meta. Meta says interactions with its AI features can be used to improve its models, so think before handing over something confidential permissions and privacy covers what is and is not included.
Four things no prompt will fix
These come up constantly in the beta, and no amount of rephrasing helps they are limits of the product rather than of your wording.
“Move these files for me”
Meta AI can look at a window and reply. It cannot click, drag, save or send. That is Muse, a separate app with a separate rollout.
“Open the file in my Documents folder”
There is no file system access. Attach the window the file is open in, or upload it in the chat nothing else reaches it.
“Look at my other screen”
Attachment is one window at a time and you pick it. It is not watching the desktop, which is the point. How it works.
“What is the current price of…”
Treat time-sensitive answers as a starting point, not a source, and confirm anything you are going to act or spend money on.
What to do when a prompt does not land
In order of how often each one turns out to be the problem.
Check something is actually attached
“This” with nothing attached produces a confident answer about nothing. The panel shows a chip when a window is handed over. If the chip never appears.
Split the prompt in two
Three asks in one sentence reliably loses one. Two turns is more reliable than one long paragraph, and faster to correct.
Say what was wrong, do not start over
“Too formal, and you dropped the deadline” beats rewriting the prompt. The thread already holds the context you would be retyping.
Name the thing you do not want
Six words of negative constraint change more than any amount of positive description. It is the step almost everyone skips.
Dictate the messy version
When the background matters more than the phrasing, speak it. You will give it three times the context without noticing. Dictation.
Ask for the working, then check it
Especially on anything with numbers or dates in it. The model got more literal in 1.3 and slightly weaker on long documents.
Prompting questions
Where do I type these prompts on a Mac?
Press ⌥Space for Quick Invoke over whatever app you are in, or open the main window from the Dock. Quick Invoke suits one-off questions; the main window is better when you want to keep the thread. Every shortcut.
Do I need to attach a window for all of these?
No thirteen of the thirty assume it, and those are marked. The rest work anywhere. Attaching is one window at a time, chosen by you, and needs Screen Recording permission the first time. How to attach one.
Why does it ignore part of my prompt?
Usually because the prompt asked for several things at once. Splitting into two turns is more reliable than lengthening it. The current model also reads instructions fairly literally, so an aside can be taken as a request. About the model.
Does being polite make the answer better?
There is no evidence it changes quality. What does is naming a constraint, a format and an audience. Politeness costs nothing either way, so write however feels natural to you.
Can I dictate a prompt instead of typing it?
Yes, and it suits the long rambling ones describing a situation out loud produces far more context than you would bother typing. Dictation types into whichever app has focus, so check the cursor is in the chat box first. Dictation guide.
Will these prompts work in the browser version?
Seventeen of the thirty will. The thirteen that depend on attaching a window need the desktop app, because a web page cannot see the window behind it. Desktop vs web.
Can I save a prompt to reuse it?
There is no prompt library in the 1.0 beta. A note file, a text expander or a clippings app is the usual workaround, and has the advantage of working across every assistant you might try. What else is missing.
How long should a prompt be?
Long enough to carry the constraint and the context, short enough that every sentence is doing work. Two or three sentences covers nearly everything on this page. Length is not the variable that matters specificity is.
Can I ask it to do the task rather than describe it?
Not in Meta AI, which reads and replies but does not act on your Mac. Taking actions is what Muse is for, and it is a separate download with its own permissions. The difference in full.
Are my prompts used to train Meta's models?
Meta says interactions with its AI features can be used to improve its models. If that matters for a particular piece of text, do not paste it and do not attach the window it is in. Permissions and privacy has the detail.
Pick one, attach a window, see what happens.
The thirteen marked prompts are the ones that will tell you whether the desktop app earns its place on your Mac.
Independent guide not affiliated with Meta. Prompts are suggestions, not commands: nothing here unlocks behaviour the app does not already have. Think before attaching anything confidential.