Best AI prompts for virtual assistants

Why virtual assistants are using AI in 2026

Three clients, three inboxes, three completely different ways of saying no. That's the part of the job the listing never mentions. A virtual assistant isn't writing as themselves, they're writing as whoever hired them, and switching register four times before lunch wears people down faster than the volume ever does.

ChatGPT and Claude help with that specific friction, and not in the way most articles claim. They don't decide anything. What they do is take the fourteenth polite decline of the week off your hands so you still have judgment left for the message that genuinely needs it.

The catch is the same one every remote assistant already knows. A model doesn't know your client, so left alone it writes generic corporate English and every client ends up sounding like the same mid-sized software company. Everything below assumes you're feeding it real context: how this person actually writes, what they've already agreed to, and what they'd never say.

The inbox, written in someone else's voice

Voice matching is the whole game, and the tone adjustment prompt is where it happens. Paste three emails your client genuinely sent, say that's the voice, then hand it the blunt two-line answer they fired back at you and ask for the sendable version. Describing a voice as warm or professional gets you nothing, because every client believes they're both. Samples carry what you'd never think to specify.

Most of the queue after that is chasing. The follow-up email prompt drafts the third nudge to a contact who's gone quiet without the passive-aggression that creeps in by attempt three, and the reminder email prompt handles the deposit, the document, or the signature you need before anything else can move. Scheduling has its own rhythm: the meeting request email prompt proposes times on your client's behalf without giving away more of their calendar than they'd like.

Two smaller ones do real work. The introduction email prompt writes the double opt-in connection between two of your client's contacts, which is a message people get wrong by making it about the introducer. And the professional out of office prompt covers the client who tells you they're leaving for Portugal ninety minutes before the flight.

When something slips, and something always slips, the apology email prompt drafts the double-booking or missed-deadline note that owns the error plainly and moves to the fix. Send that one from your own judgment about who's actually annoyed.

Calendars, meetings, and what comes out of them

Meetings generate more admin than they contain. The meeting agenda prompt turns your client's scattered notes into a structure the attendees can prepare against, which is the difference between a call that decides something and a call that gets repeated next week. Afterwards, the meeting summary prompt pulls decisions and owners out of a messy transcript, and pulling out the owners is the part that saves you the follow-up thread.

Then there's the queue itself. The task prioritization prompt sorts a morning where two clients both marked something urgent, which is a weekly event once you're past one client. It won't make the call for you, but laying the conflict out in writing usually makes obvious which client needs a heads-up.

Reporting, or proving the week happened

Invisible work is the structural problem of remote assisting. Nobody sees the eleven small things that stopped a problem, so a quiet week can look like an idle one.

The weekly review prompt turns your own scrappy log into the client-facing summary that makes the retainer feel earned, and it's worth sending even when nobody asked for it. For anything running across weeks, an office move, a hiring round, a site launch, the project status report prompt gives your client the state of it in a format they can forward without editing.

Documents, briefs, and reading nobody has time for

The summarize long document prompt is the one VAs use most after the email prompts. A forty-page supplier agreement, a research report, a competitor's deck, all of it becomes the brief your client reads instead. Give it a standing instruction to flag anything that looks like a commitment or a date, since those are the lines they'll ask about.

Cleanup work splits in two. The proofreading prompt catches the things that embarrass a client in a document going to their board, while the grammar check prompt is the fast pass before a routine email leaves. When you're asked to compare venues, tools, or suppliers, the pros and cons prompt organizes what you found into something your client can decide from in a minute. Present the comparison. Let them pick.

Procedures: the deliverable that outlasts the task

Writing down how you do something is the highest-value thing a VA produces, and it's always the thing that gets postponed. The SOP prompt converts a process you've run twenty times into a document another person could follow, which matters twice over: it's what you hand a client's new hire, and it's what lets you take a holiday without the work stopping.

The onboarding checklist prompt builds the intake list for your own new clients, the one that gets you passwords, calendar access, voice samples, and preferences in week one instead of extracting them painfully over two months. And the FAQ generator prompt drafts the canned responses for questions that hit your client's inbox weekly, so the fifth identical enquiry costs you a paste instead of a rewrite.

Your client's customers, handled at arm's length

Plenty of assisting work faces outward. The customer onboarding email prompt writes the welcome sequence a client's new customer receives, and the customer review response prompt drafts replies to public feedback, which is delicate precisely because it's posted under your client's name and not yours. Send those for approval the first few times until you know their line.

Facing the other way, the complaint email prompt writes the firm, factual message to a supplier who shipped late or billed twice. Drafting that one with help is genuinely easier than writing it while irritated on someone else's behalf.

Social and content scheduling

Content is the work that quietly expands to fill a retainer. The social media caption prompt drafts posts in the client's voice rather than the default LinkedIn register, and the content calendar prompt maps a month so you're scheduling rather than inventing something at 9am. For clients who publish audio, the podcast show notes prompt turns an episode into notes, timestamps, and pull quotes, which is a task that used to eat an hour per episode.

Finding clients and staying findable

The part of the job nobody assigns you is filling the next slot, and most VAs are worse at selling themselves than at any client task.

The cold email prompt drafts the pitch to a business showing the symptoms of an overloaded founder, led by what they get rather than your service list. Once someone bites, the business proposal prompt structures the retainer: scope, hours, response times, and what sits outside the agreement, which is the section that prevents scope creep six months later.

Inbound matters as much. The LinkedIn profile summary prompt rewrites the section clients actually read before messaging, and the LinkedIn headline prompt fixes the line that determines whether you appear in their search at all. When you're applying rather than pitching, the job application email prompt tailors the note to the listing, which beats the template every other applicant sent.

Which AI tool for which VA task

Match the tool to the task, and see the ChatGPT vs Claude vs Gemini guide for the fuller comparison. Gemini is worth a look when a client's whole operation already lives in Google Drive.

Where AI falls short for virtual assistants

Three failure modes matter more in assisting than in most jobs, because the output goes out under someone else's name.

It invents specifics. Ask for a reply about an order, a booking, or an invoice and a model will fill the gaps it has no way to know: a delivery date, a reference number, a policy that sounds plausible. Anthropic's own guidance on reducing hallucinations amounts to giving the model the source material and letting it say it doesn't know. In a client's inbox, an invented detail is a commitment they now have to honour.

It doesn't know what your client already agreed. The model sees the message in front of it, not the phone call last Tuesday or the discount promised in March. Anything referencing history has to carry that history in the prompt or come back to you before it sends.

And confidentiality is yours, not the tool's. You're handling other people's data, so the tier you're on and what it does with conversations is worth knowing rather than assuming, and both providers publish the detail, including Anthropic's policy on training data. Placeholder names cost you five seconds.

Where to start

Pick the email you retype most often, probably a scheduling reply or a polite decline, and build one reusable prompt for it with three real samples of your client's writing pasted in. Save it per client. That single prompt covers more of a working week than any other on this page.

After that, the five variables that make a prompt work apply directly to voice matching, and both OpenAI's prompt guide and Anthropic's land on the same point: specificity beats instruction. The signal you've got it right isn't that the draft reads well. It's that your client sends it without changing a word, and doesn't mention it.