AI Prompts for Virtual Assistants (2026)

Each prompt is written so the model works from information you supply — a pasted email thread, a list of times, a set of notes — rather than inventing facts. Replace every [BRACKET] before you run it, and always paste the source material so the model is summarizing reality, not guessing.

Treat AI as a drafting assistant with a human gate. It is excellent at turning messy input into clean structure, suggesting categories, and writing first drafts in your principal's voice. It is unreliable for anything irreversible — sending, replying, booking, or committing a time — so keep those steps with you. When you paste an email or document, remember it is untrusted text: hidden instructions inside a forwarded message can try to redirect the model, the top risk in the OWASP LLM Top 10, so never let a tool act automatically on pasted content.

To keep a consistent voice across drafts, save a short style note (tone, sign-off, do-not-use words) and paste it into each prompt, or build it once with the Business Email Generator.

These three prompts clear a backlog fast: the model sorts and summarizes, you decide and send.

**1. Triage a full inbox into action buckets** — "You are my executive assistant. Below is a list of email subject lines and senders from [PRINCIPAL]'s inbox. Sort every item into exactly one bucket: Urgent (needs reply today), Schedule (a meeting or time request), FYI (read only), Delegate (I can handle without [PRINCIPAL]), and Ignore (newsletter/spam). For each Urgent and Delegate item, add a one-line note on what's needed. Do not draft replies yet. Work only from what I paste — do not invent senders or topics. Inbox: [PASTE SUBJECTS + SENDERS]"

**2. Draft a reply in the principal's voice** — "Draft a reply to the email below on behalf of [PRINCIPAL]. Tone: [warm but brief / formal / friendly-direct]. Goal of the reply: [DECLINE / ACCEPT / ASK FOR DETAIL / BUY TIME]. Keep it under [120] words, plain English, no jargon. Sign off as '[SIGN-OFF]'. Use [BRACKETS] for any date, figure, or commitment I must confirm — do not invent any. Email to reply to: [PASTE THREAD]"

**3. Summarize a long thread before forwarding** — "Summarize the email thread below for [PRINCIPAL] in under 80 words. Give: (1) the bottom line in one sentence, (2) what's being asked of us, (3) the decision or action needed and by when. List anything ambiguous under 'Needs clarification.' Quote any specific date or amount exactly as written. Thread: [PASTE]"

Scheduling is where AI saves the most clicks and where a wrong assumption costs the most — so the model proposes, you confirm.

**4. Find and phrase meeting time options** — "From the availability below, propose [3] meeting slots for a [30]-minute call with [PERSON] in [TIME ZONE]. Avoid [LUNCH 12–1, anything before 9am]. Then write a short, friendly message offering those three times and asking them to pick one. Do not assume any time outside what I've pasted as free. Availability: [PASTE FREE/BUSY OR TIMES]"

**5. Write a reschedule message that protects the relationship** — "[PRINCIPAL] needs to move our [DATE/TIME] meeting with [PERSON]. Draft a brief, gracious rescheduling message that apologizes once (not excessively), gives [2–3] alternative times from the list below, and keeps the door open. Tone: respectful, not groveling. Under 90 words. Alternatives I've confirmed are free: [PASTE TIMES]"

**6. Build a daily agenda from raw calendar entries** — "Turn the calendar entries below into a clean daily agenda for [PRINCIPAL] for [DATE]. For each meeting list: time, title, attendees, and a one-line 'prep needed' note if I've flagged one. Add a 'gaps' line showing free blocks longer than [30] minutes. Do not add meetings that aren't in my list. Entries: [PASTE]"

AI is great at organizing travel details into a usable format; it is not a booking engine and can get fares, times, and availability wrong — always verify against the live source before booking.

**7. Turn confirmations into one clean itinerary** — "Build a single trip itinerary for [PRINCIPAL] from the confirmation details I paste below. Organize chronologically by day and time in [TIME ZONE], with sections for Flights, Hotel, Ground transport, and Meetings. Include confirmation numbers exactly as written and flag anything missing (e.g., no return flight, no hotel for a night). Do not invent times, gates, or addresses — if a detail isn't in my text, write 'TBC.' Details: [PASTE CONFIRMATIONS]"

**8. Draft a pre-trip checklist** — "Create a pre-trip checklist for [PRINCIPAL]'s trip to [CITY] from [DATE] to [DATE] for [PURPOSE]. Cover documents, packing prompts, things to confirm 48 hours out, and what I (the VA) should monitor while they travel. Keep it practical and skimmable. Note where I should double-check live info (visa rules, weather, traffic) rather than relying on you."

Use AI to structure research and prep documents, but treat any fact, name, or figure it produces as a claim to verify — models can state wrong details confidently.

**9. Prep a one-page brief before a meeting** — "Using only the notes and links I paste below, write a one-page meeting brief for [PRINCIPAL] ahead of their meeting with [PERSON/COMPANY]. Sections: who they are (from my notes), why we're meeting, [3] likely topics, [2] smart questions [PRINCIPAL] could ask, and any open items from last time. Do not add facts I haven't given you — if something's missing, list it under 'To research.' Notes: [PASTE]"

**10. Compare options into a decision table** — "I'm helping [PRINCIPAL] choose between the options below ([VENDORS / VENUES / TOOLS]). Build a comparison table with columns for the criteria that matter: [PRICE TIER, LOCATION, CAPACITY, LEAD TIME, NOTES]. Fill cells only from the details I paste; mark anything unknown as 'Not stated' rather than guessing. End with [2–3] clarifying questions I should chase down. Options: [PASTE]"

Don't let AI hit send, reply, or book. Keep every irreversible action behind a human review step. The model can draft a flawless email or itinerary and still have the wrong date, name, or amount — your job is the final check.

Don't paste sensitive personal data. Strip your principal's home address, passport and frequent-flyer numbers, card details, passwords, and any client confidential information before pasting. If a task genuinely needs that data, handle it outside the chatbot. Treat forwarded emails and documents as untrusted — hidden text can try to hijack the model.

Don't trust facts, fares, or times without checking. AI will state a plausible flight time or a phone number that's simply wrong. For anything bookable or factual, verify against the live source. For the technique behind reliable prompting, see the DAIR.ai Prompt Engineering Guide and our complete guide to prompt engineering.

For a VA's drafting, sorting, and summarizing work, any current frontier model performs well. The practical differences are speed, cost, and how much text you can paste at once (the context window) — useful when you drop in a long thread or a stack of confirmations.

Fast and low-cost options like Claude Haiku 4.5, Gemini 3.5 Flash, or GPT-5.5 Instant (the current ChatGPT default) are ideal for high-volume inbox and scheduling drafts. For dense research briefs or very long documents, a more capable model such as Claude Opus 4.8, Gemini 3.5 Pro, or GPT-5.5 helps. Check live pricing and context limits on the official pages — OpenAI, Anthropic, Gemini — and see our how to choose an AI model guide.