Manus AI Playbook: Modes, Wide Research, Skills and Prompts

Manus AI works as an agent, not a chatbot: each task gets its own cloud virtual machine, and you steer it mainly by picking a mode. Manus 2.0, launched in September 2026, offers four: Manus 2.0 Max for complex jobs, Manus 2.0 for most tasks, Manus 2.0 Lite for everyday work and small edits, and Manus Flex (beta) for your own model key. Free accounts get 2.0 Lite in Agent mode, plus Chat; Manus does not say which plans can use Flex. Wide Research, the parallel pattern, starts by itself on paid plans and caps each sub-task at 50 credits. The habit that matters most for credits is plain. Ask precisely, split big jobs, and keep quick questions in Chat.

Pick the mode before you type

The mode menu decides both the depth of the work and the spend. Manus's website-builder docs make the split concrete: Max for full-stack apps with a backend, database, login or payments, 2.0 for most websites, Lite for small edits. Free users get Manus 2.0 Lite in Agent mode only, while Pro users get all three ^[1]. The 300 daily refresh credits every account receives work only in Lite, on free and paid plans alike.

Chat mode and Agent mode are a separate choice. Agent mode plans and finishes work on its own: websites, slides, videos and now mobile apps. Chat mode reads images, PDFs and text files and searches the web, and Manus recommends it for simple questions to save credits. It is not free, though. Both modes now draw on the same balance. Manus enters Chat only when you pick it from the plus or slash menu, never on its own. You can move a conversation from Chat to Agent. You cannot move it back.

Coming from the 1.6 generation

Older guides still describe the previous generations. Version 1.5 arrived in October 2025, and version 1.6 followed in December 2025 with a flagship Max agent next to a Lite agent. The split survives in the 2.0 menu, with Lite open to free accounts and the stronger modes reserved for subscribers, so old Max habits belong in 2.0 Max and old Lite habits in 2.0 Lite. Any guide quoting 1.6 limits or the old Basic and Plus plan names describes a product that has since changed.

What runs behind a single task

Every task runs in its own isolated cloud virtual machine with internet access, a file system and a browser, with root access inside. Manus does not publish the machine's specs because it adjusts them dynamically. An idle sandbox sleeps, and one that keeps sleeping may be recycled after 7 days on Free or 21 days on Pro. Manus's pages disagree on what survives a sandbox reset. Download anything you need to keep. Waking a sleeping sandbox (for example to reopen a preview) by sending it a message also spends credits.

Manus 2.0 runs on Cascade, an in-house agent harness that starts each project light and adds specialized capabilities only when the work calls for them. Manus says that in one tested configuration Cascade used 23.2% fewer tokens, finished 28.2% faster and cost 32% less than the previous system, and it notes this is not an overall average. Treat it as one vendor benchmark, not a forecast for your tasks. The founder also says 2.0 still has rough edges.

For media, the agent calls built-in models that bill Manus credits and need no provider account. Video runs on Seedance 2.5, Seedance 2.0, Veo 3.1, Gemini Omni Flash or MiniMax H3; only Gemini Omni Flash edits an existing video, and Veo output on Manus is 16:9 landscape only. Images default to GPT Image 2.5, the one model that produces transparent PNGs, beside three Nano Banana variants. Audio covers Lyria 3.5 songs of up to about three minutes, Gemini 3.8 Flash TTS and ElevenLabs sound effects. Video generation requires a paid plan.

Manus Flex: your key, Manus's agent

Manus Flex launched in September 2026 and is still in beta. You connect a key from OpenAI, Anthropic, Google AI Studio, xAI, Meta, OpenRouter, Fireworks or Modal, then choose the model and an effort level ^[2]. Your provider bills the inference and Manus covers the sandbox, while images, videos, Cloud Computer and website hosting still spend Manus credits. Built-in media models never switch to your key. There is no custom endpoint or self-hosted model field, and if your key runs dry, tasks can stop or fail.

Wide Research: the parallel mode you cannot switch on

Wide Research splits a job into independent sub-tasks and hands each to a full Manus instance with its own virtual machine and fresh context. The sub-agents never talk to each other; the main agent merges their results into a table, report or dataset. Per the help center, it starts only when the AI decides a task can be split, cannot be started manually, runs for paid users only, handles 20 sub-tasks at a time and caps each at 50 credits ^[3]. In 2.0 Max, every sub-agent uses the more advanced architecture.

Since there is no button, the lever is the shape of your request. Manus says the pattern suits many similar items and advises against it for fewer than 10 items, steps that depend on each other, a single deep dive or real-time research. Its docs say it has been tested up to 250 items, and it also handles batch image editing and data extraction. Talk of hundreds of agents stays Manus's claim. Analysts will find data sources and citation caveats in our guide to Manus AI for researchers.

Making Manus repeat work without re-explaining it

Plan Mode is the brake. Open to all users and started manually (type / and choose Plan on the web, or tap + on mobile), it writes an editable Markdown plan and builds nothing until you confirm or dismiss it. You can switch it on mid-task. Branch covers the opposite need: it opens a new task from any reply, carrying the history and files up to that point, and leaves the original untouched. Only the task owner can branch, up to 10 times a minute, and website projects cannot be branched.

Projects hold a master instruction and a knowledge base that every new task inside them inherits, on all tiers and with no cap on their number. Instruction edits reach running tasks at the next message; new files reach only later tasks. Skills follow the Agent Skills open standard introduced by Anthropic. You can package a finished task as a Skill, upload a zip or .skill file, or import one from GitHub or the official library, then call it with a slash command. Community Skills can contain shell commands. So read them before use.

Automations, which absorbed Scheduled Tasks, come in three types: a schedule, a triggered task and an advanced automation (Beta) that turns a plain-language request into a workflow. Schedules run at most every five minutes. Triggers come from Gmail, Outlook, Mail Manus, calendars, GitHub, Shopify, Slack, Notion, Meta Ads, Google Ads, Instagram, RSS or a webhook. GitHub, Shopify, webhook and Mail Manus events arrive instantly; the rest are checked at intervals. A trigger can pause after the same failure repeats. Run a test first.

Each automation also sets whether work that sends or publishes needs your approval. The Skip confirmations option lets a trusted schedule send, publish or post without asking, so grant it last. Scheduled tasks keep running in the cloud with your device closed. Routines that need local files or apps belong in Manus Studio (the desktop app, formerly Manus Desktop), and developers wiring Manus into their own stack should start with our Manus API guide.

Prompts that waste fewer credits, and answers to double-check

Manus cannot estimate a task's credit cost before it starts ^[4], and it says any figure the agent itself quotes should be treated as a hallucination. Cost grows with complexity, duration and context length. So state the topic, scope and output format up front, since a specific request spends fewer credits than a vague one. Split complex work into smaller tasks. Keep simple questions in Chat. Afterwards, check Settings, Usage to learn what similar jobs cost, and remember that unused monthly plan credits do not roll over.

Manus also publishes where its agent goes wrong: promises about credits, admissions of system-wide problems, pledges to do impossible things, fabricated content such as invented papers or citations, and logical contradictions. It warns that output can sound confident and still be inaccurate, and it shows an informational High-impact topic notice in health, finance and legal conversations without blocking them. The Terms put the review burden on you. Ask for sources, open each one, and verify every number before the work leaves your desk.

The bottom line

Manus rewards users who decide before they type: Lite for small edits and the daily refresh credits, 2.0 for most work, Max for complex builds, Chat for quick questions. Wide Research is its most distinctive pattern, but it fires only when a request is shaped as many similar items, and only on paid plans. Plan Mode, Projects, Skills and Automations turn one good task into a routine. Credits remain the unknown. Nothing estimates them in advance, so start small, read Usage after each run, and check every claim the agent makes.