Meta AI stopped being a novelty somewhere around the middle of 2025. It now sits inside WhatsApp, Instagram DMs, Messenger, a standalone app, a browser tab, a pair of Ray-Ban glasses and a desktop client. That reach is exactly why so many people bounce off it: they poke the blue circle, ask one vague question, get one vague answer, and never open it again.
What follows is the setup that actually sticks. Six moves, in order, with the exact prompts I’d hand to someone who has ten minutes and low patience.
Find the surface that fits your day
Meta AI behaves slightly differently depending on where you open it, and picking the right door saves more time than any prompt trick.
- WhatsApp — the circle above your chat list, or type @MetaAI inside a group. Best for plans, translations and making sense of long forwarded threads.
- Instagram and Messenger DMs — search “Meta AI” in the message bar. Handy for captions and comment replies without leaving the app.
- The standalone app — worth installing properly. It climbed to No. 5 on the US App Store after the Muse Spark launch reshaped how the assistant handles voice, and it carries the fullest feature set.
- Desktop — meta.ai in any browser, or the Mac client, which is designed around controlling other apps rather than just chatting. The Mac app wants you to talk to your apps, and that framing changes what you should ask it for.
- Ray-Ban and Oakley Meta glasses — the only surface where “look at this and tell me” is the entire point.
If you only pick one, pick the app. If you live in WhatsApp, stay there. Splitting your usage across four surfaces is how you end up with no muscle memory at all.
Write the first prompt like a brief, not a search
The biggest difference between people who find Meta AI useful and people who find it flat is the shape of the request. A search query and a prompt are not the same object.
Compare these two. “ideas for dinner” gets you nine generic dishes you’ll scroll past. “I have 20 minutes, one pan, chicken thighs, a lemon and half a cabbage. Give me two options, one of them spicy. Keep each under 100 words.” gets you something you can actually cook tonight.
Name the format you want
Meta AI defaults to cheerful paragraphs. If you want a table, a numbered list, a 40-word summary or a message you can paste straight into a chat, say so. “Reply as a WhatsApp message I can forward” is a genuinely useful instruction and it costs four words.
Paste the raw material instead of describing it
Long-press a thread and hand it over. Meta AI is far better at working with text you give it than at guessing context you didn’t. This is where the WhatsApp version earns its place: forwarding a 60-message group thread and asking “what did we actually decide?” takes about nine seconds.
Five prompts worth stealing
Concrete beats clever. These are the ones that survive contact with a real week.
- Trip logistics: “Here are 14 messages about a weekend in Lisbon. Pull out the dates, who’s confirmed, and the three things still undecided. Bullet points only.”
- Caption drafting: “Write three Instagram captions for a photo of a rainy street at night. One funny, one dry, one under six words. No hashtags.”
- Glasses lookup: “What plant is this, is it safe around a dog, and how often do I water it?” — asked while pointing at the leaf, which is the one thing a phone can’t easily do.
- Inbox triage: “Summarise this email in two lines, then draft a polite reply declining the meeting and suggesting the first week of March instead.”
- Study helper: “Explain how compound interest works to a 15-year-old, then give me five practice questions with the answers hidden at the bottom.”
Notice the pattern running through all five: context in, format specified, length capped. Once that becomes automatic you stop writing instructions and start just talking.
Let it take actions, not just answer questions
The 2025 rebuild pushed Meta AI away from pure chat and toward an agent that does things on your behalf. Some of that shipped — reminders, calendar awareness, cross-app requests on desktop — and some is still arriving, which is the wider story behind Meta’s shift toward agentic behaviour.
The practical version today is simple: ask it to produce something rather than describe it. A drafted message. A checklist. A set of notes renamed and reordered. On a Mac, the interesting behaviour is the handoff — you describe the outcome (“move the Q3 invoices into a folder and email me the list”) and the assistant works across your windows. It fails in boring, obvious ways, but the ten minutes you spend poking at it now means you’re not learning it cold next year.
Generate images with realistic expectations
Imagine, Meta’s image tool, is fast and cheerful and produces people with too many fingers about one time in eight. The Muse experiments are stranger still — capable enough to be genuinely unsettling, which tells you the models are running ahead of the guardrails.
What it’s good at: mood boards, textures, backgrounds, thumbnail concepts, “show me four versions of this idea”. What it’s bad at: anything with text in the image, anything photorealistic that needs to be believable, and anything involving a real person’s face. If a generation looks off, change one variable at a time instead of rewriting the whole prompt.
Set the boundaries on day one
Two settings are worth changing before you rely on Meta AI for anything that matters.
Memory and data. Meta AI can remember details across conversations. Open the app settings and check what’s stored, and whether your glasses captures are feeding model training. Turn off whatever you don’t actively want.
Group participation. Inside a WhatsApp group, Meta AI only reads messages when someone @-mentions it. That’s by design, and it’s worth knowing so nobody in the group assumes otherwise.
The wider context matters too. Meta’s own ad systems have promoted apps built to strip clothing from photos of real teenagers, which is a reminder that “the assistant is helpful” and “the platform’s incentives match mine” are two separate claims. Use it, but don’t hand over the family photo archive in week one.
Build a two-minute loop that pays off
Habits beat feature tours. Here’s a loop to run for a fortnight:
Monday. Forward your messiest group thread and ask for the three open decisions. Tuesday. Dictate a rambling voice note and ask for a three-line summary you can paste into Slack. Wednesday. Photograph something you don’t recognise using the glasses and get an ID. Thursday. Draft one reply you’ve been avoiding, then edit it in your own voice. Friday. Ask what you queried most this week and pick a single thing to automate properly.
Two minutes a day, two weeks, and you’ll know exactly which surface earns its place in your routine and which one you can quietly delete. That’s a more useful outcome than any review can hand you, because the answer depends entirely on the shape of your own week — and now you’ll have the data to see it.