Muse Spark Tutorial: Five Short Lessons for Using It at Work

Muse Spark is Meta’s AI model, and it runs the Meta AI app and meta.ai. If you already use AI at work, you do not need a new way of thinking to get good results from it. You need five habits, practised on real tasks. This tutorial gives you one lesson per habit, each with a weak prompt, a better one, and a note on why the second works.

One naming point before we start, because it trips people up. Muse Spark is the model. Muse is Meta’s separate agent app, built on that model. If you want the longer background first, what Muse Spark is and where it came from covers it. If you have not opened it yet, how to use Muse Spark in the Meta AI app, on meta.ai or through the API covers getting in.

Lesson 1: Pick Instant, Thinking or Contemplating

Meta’s launch post describes Instant and Thinking modes in chat, plus a Contemplating mode that has several agents reason in parallel. Contemplating was rolling out gradually, so you may not see it yet. The point of the lesson is simple: match the mode to the job, not to how important the job feels.

• —Instant — quick rewrites, short replies, a first draft of something you will edit anyway.
• —Thinking — anything with steps: a plan, a comparison, a sum you need to be right.
• —Contemplating — a hard question where you would happily wait for a more careful answer, if your account has it.

Before, in Instant: "Should we move our team offsite from March to May?"

After, in Thinking: "We have 14 people and a budget of £6,000. March is our busiest sales month. May clashes with two people’s leave. Compare the two dates on cost, attendance and workload, then recommend one and say what would change your mind."

The first prompt in a fast mode gets a fast, generic answer. The second gives a slower mode something to actually reason about. A slower mode does not rescue a vague question, so lesson 2 matters more than lesson 1.

Lesson 2: Write a brief with goal, context and format

Most weak output comes from a weak request. The fix is the one OpenAI’s prompt guide describes, and Anthropic’s prompt engineering overview says much the same. Say what you want, give the facts only you have, and say what shape the answer should take. It works the same way on Muse Spark as anywhere else.

Before: "Write an update about the project delay."

After: "Goal: tell our client the website launch moves from 3 October to 17 October. Context: the payment provider changed its checkout rules and we need two weeks to retest. The client is patient but has a campaign booked. Format: an email under 150 words, plain and direct, one apology, and a clear next step. No bullet points."

Goal, context, format. Each line removes a guess the model would otherwise make for you, and its guesses sound as confident as its facts. Meta says Muse Spark 1.3 asks clarifying questions when a prompt is ambiguous, which helps, but answering those questions up front saves a round. The full method, with more examples, is in how to write a prompt that works on the first try.

Lesson 3: Give it a file or an image

Pasting a summary of a document loses detail. Giving the model the document itself does not. Meta says Muse Spark reads images, video and PDFs, and a hands-on review found the chat could find and count objects in an image, a tool known as visual grounding. That is useful for dull, real work: counting stock on a shelf photo, checking a slide against a brief, pulling figures out of a scanned invoice.

Before, with a photo attached: "What’s in this picture?"

After, with the same photo attached: "This is a photo of our stockroom shelf. Count the blue boxes and the white boxes separately. If any box is partly hidden, count it and tell me how many were partly hidden. Give me the result as two lines, then one sentence on anything you were unsure about."

The after prompt names what to count, says what to do with the awkward cases, and asks the model to flag its own doubt. Then spot-check it. Count one row yourself. If the model is wrong on the row you counted, it is wrong elsewhere too. For longer documents, the same idea sits behind getting an AI report generator to produce a draft you can trust: give it the source, and ask it to point to where each claim came from.

Lesson 4: Have it search and cite, then check the citations

Muse Spark can search the web. The same hands-on review found web search and page opening among the tools in chat, and Meta lists built-in search with citations for developers. Search makes answers more current. It does not make them correct.

Before: "What are the new rules on AI in hiring?"

After: "Search for official guidance published in 2026 on using AI tools to screen job applicants in the UK. Only use government or regulator websites. For each point, give the source link and quote the sentence that supports it. If you cannot find an official source for something, say so rather than filling the gap."

Now do the part people skip. Open every link. Find the quoted sentence on the page. Check the date. A citation can exist and still not say what the answer claims it says. A study led by the EBU and the BBC found that AI assistants often misrepresent news content. The wider research on why language models produce fluent, false statements explains why this will not simply go away.

If the topic is outside your expertise and you cannot judge the source yourself, how to check an AI answer when you are not the expert gives you a routine for it.

Lesson 5: Hand a multi-step job to Muse, with approval

Chat gives you answers. Muse, the agent app, does tasks. Meta says Muse can send emails, book travel, fill out forms and make purchases, and keeps working after you close the app. As of September 2026 it is rolling out in the US only. Setting up Muse, Meta’s agent app walks through the first run.

The skill here is not a clever prompt. It is deciding how much access to give and where the agent must stop and ask. Meta says you choose which apps Muse connects to and how much access it gets, such as read-only email versus sending, and that it asks for approval before sending an email or making a purchase.

Before: "Sort out the supplier meeting next week."

After: "Find a 45-minute slot next Tuesday or Wednesday when I’m free, between 10am and 4pm. Draft an email to Priya at our packaging supplier proposing two of those times, and mention we want to review the new carton prices. Show me the draft before sending. Do not accept or move any other meetings."

The after prompt names the steps, the limits, and the checkpoint. It also says what not to touch, which matters more with an agent than with chat, because an agent’s mistake is an action, not a paragraph. Read the draft when it asks. Approval only protects you if you actually read what you approve.

Start with small, low-stakes jobs and read-only access, and widen it once you have seen how it behaves. If you are unsure which jobs are worth handing over at all, find the repetitive part of your week first. The privacy side, including what Meta says it does with your data, is covered in whether Meta’s Muse is safe to use.

Practice exercise

Take one real task from this week. Something you would normally do by hand, like a status update, a short summary or a comparison. Then work through these steps.

1. 01Write the lazy one-line prompt you would normally type. Save the answer.
2. 02Rewrite it with a goal, the context only you know, and a format. Choose Instant or Thinking on purpose.
3. 03Attach the source file or image if there is one, and ask the model to say where each claim came from.
4. 04If it needs current facts, ask it to search and cite. Open two of the links and check the quoted sentence is really there.
5. 05Compare the two answers. Write down what the second prompt added that made the difference.

That last step is the one that builds the habit. After a few rounds, you will write the better prompt first. You will also start to notice the kinds of task where no prompt helps, which what AI is actually bad at lists plainly.

A good brief gets you a better draft. Checking the draft is still your job.

Coursium teaches this kind of practical AI skill for work, with short lessons and practice tasks on your iPhone. If you want to keep building the habit, start with Coursium.