How to Use Manus: Your First Autonomous Agent Task

Most people meet Manus expecting a chatbot and are surprised when it opens a browser and starts working. Manus describes itself as an autonomous general agent that operates inside a sandboxed virtual environment with internet access, persistent file storage, and the ability to install software. That difference changes how you should talk to it.

This walkthrough covers the first hour: what the two modes are for, how to write a task Manus can finish without babysitting, what the run looks like while it happens, and the handful of mistakes that make a first task disappointing. Everything here comes from Manus's own docs and help centre, plus independent hands-on reviews, all linked at the end.

What do you need before you start?

You need an account at manus.im and a task that has more than one step. Nothing is installed by default, because Manus runs in the cloud in its own environment. Two optional add-ons matter later but not on day one: the Browser Operator extension and any tool connectors you want it to reach.

It also helps to know that Manus bills in credits rather than messages. Their help centre states that credits are deducted only while the agent is actively processing a task, and that simply asking questions in chat mode without triggering an agent workflow does not consume credits. Cost is driven by task complexity, meaning the tokens and compute used to plan and execute.

Chat mode or agent mode: which one do you want?

Use chat when you want an answer, and agent when you want a result. That is the whole distinction, and getting it wrong is the single most common reason a first Manus session feels underwhelming.

Manus's own guidance for beginners is blunt about this: if a regular chatbot could answer your request in one response, it is not an agent task. Their examples of agent work involve taking actions across steps, tracking progress, making decisions, following through, and producing structured output you can reuse. Phrases like take ownership, track this over time, and follow up when needed are the tells that you are actually using the agent.

Notice that every task on the right names a deliverable. Manus is built to hand back a finished artefact, so telling it what the artefact is does most of the work of directing it.

How do you write a first task Manus can actually finish?

Give it a goal, a boundary, and an output format. The goal is what done looks like, the boundary keeps it from wandering, and the output format tells it when to stop researching and start writing.

Manus's own credit-optimisation advice points the same direction. They recommend grouping related questions into one task rather than firing them separately, stating your expected outcome once instead of issuing several vague instructions, and reading what came back carefully before resubmitting a variation. Vague prompts cost more because the agent explores more.

Running your first task

1. Pick a task with a real deliverable — Choose something that ends in a file or a structured list: a comparison table, a short research memo, a cleaned dataset. Avoid open-ended requests like organise my work, which have no finish line.
2. Write the goal, the scope and the format in one message — State what you want, the limits (how many items, which time window, which sources or regions), and what the output should be. Manus's documentation for scheduled work makes the same point about precision, preferring an exact time window over a vague one.
3. Attach anything it should not have to guess — If the task involves your data, upload it. Manus's data documentation lists PDF, CSV and Excel formats among the structured files it accepts, so give it the source rather than describing it.
4. Start the task and keep the run panel open — Manus shows its work in a panel commonly referred to as Manus's Computer. Independent reviewers describe watching it browse and being able to step in at any point, which is far more useful than reading a summary afterwards.
5. Correct it early, not at the end — If you can see it heading somewhere wrong in the first few minutes, say so in the same conversation. Manus incorporates new instructions while keeping track of what it already completed.
6. Check the output before you use it — Reviewers who have tested Manus on real research consistently report needing to spot-check the result. Treat the first draft as a strong starting point, not a finished document.

The step that people skip is the fifth one. An autonomous agent that has spent forty minutes down a wrong path cannot get that time back, and neither can you. Watching the first three or four actions costs you a minute and saves you the whole run.

How do you keep your setup from repeating itself?

Manus has a Projects feature for exactly this. Their documentation describes a project as a dedicated workspace where you define a master instruction and build a knowledge base that applies automatically to new tasks inside it.

You create one from the plus icon in the sidebar, name it, write the master instruction, and upload the files it should always know about. One behaviour worth remembering: their docs say instruction changes take effect on your next message, while file updates only apply to newly created tasks, so existing tasks keep the configuration they started with. Manus states there is no cap on how many projects you can create.

There is also a Skills system for reusable procedures. Manus documents these as modular instruction sets managed from a Skills tab in the left-hand menu, added by building one from a successful run, uploading a file or folder, choosing from their official library, or importing from GitHub, and then invoked by typing a forward slash in chat. Their docs also warn that community-contributed skills can contain executable code and should be reviewed before use.

What should you expect to go wrong?

Three things, and none of them mean you did it wrong. Knowing about them in advance is the difference between a useful first session and an abandoned account.

• Gated content stops it. MIT Technology Review's hands-on test reported repeated CAPTCHA blocks and paywalled sources it could not reach, along with sites that flagged the activity as suspicious. Anything behind a login or a paywall is a likely stall point.
• Big collection jobs come back thin. The same review found that a request for fifty fully profiled candidates returned three after three hours. Scope your first task to something it can genuinely finish.
• Accuracy needs your eyes. Independent reviewers put practical accuracy in the range where the work saves hours but still needs spot-checking, and one presentation review noted that Manus does not surface which sources it drew from.

If reading that list gave you pause, it is worth naming the trade-off honestly. Manus is a general agent, which is what makes it impressive: it will attempt almost anything, and the cost of that flexibility is that you supervise the attempt. That is a genuinely good deal for one-off, exploratory work.

It is a worse deal for the same job every week. Recurring work is exactly what we built Sistava around. Instead of writing a fresh task each time, you hire an AI Employee, tell it in plain English what it owns, and it keeps the context, the tools and the schedule itself. There is no prompt to rewrite because the responsibility, not the instruction, is the thing you handed over.

What is a good second task?

Something that repeats. Manus supports scheduled tasks, and their documentation describes setting a frequency such as daily, weekdays only, weekly, monthly, a custom pattern, or a single delayed run, then defining where the output goes. Active schedules are managed from Settings, where you can pause, edit or delete them.

Their guidance on writing these is the same as for any task, only stricter: give an exact time including the zone rather than a word like morning, and give research jobs an explicit time window such as the past twenty-four hours. Vague scheduling produces drifting results week to week.

Sources

• Manus documentation, Welcome
• Manus documentation, Projects
• Manus documentation, Manus Skills
• Manus documentation, Scheduled Tasks
• Manus documentation, Data Analysis and Visualization
• Manus blog, Beginner Prompts to Get Started with Manus
• Manus Help Center, Understanding credit consumption in the Manus desktop app
• Manus Help Center, How to optimize my credit usage
• MIT Technology Review, hands-on test of Manus
• Plus AI, Manus AI slide generator review

Start with one task that has a clear finish line, watch the first few minutes, and correct it while it is still cheap to correct. Manus rewards specific direction more than clever phrasing, and almost every disappointing first session traces back to a request a chatbot could have answered.