AI Central's walkthrough for Claude Cowork covers eight steps, and the ones that matter are not the install. Download the desktop app on a Pro plan, open the Cowork tab, pick Opus 4.6 with extended thinking on, then build a single folder called Claude-Work with four subfolders, write three context files, set standing instructions once, and force Claude to interview you before it acts. Plugins and connectors come last, on purpose.
The problem this setup is actually solving
Most people meet a new AI tool the same way. Paste in some context, correct the output, paste more context, then start from zero the next morning. AI Central frames the Cowork setup as the fix for exactly that, the move from copy-paste chaos to a workflow that remembers who you are between sessions.
Cowork is a desktop surface rather than a browser tab, and it carries its own research preview notice on the way in, with a pointer to learn how to use it safely. As of August 2026 that means this is a setup you revisit, not one you build once and forget.
Install and model choice, the two minute part
AI Central's first two steps are the only purely mechanical ones. Get the desktop app from claude.com, which needs a Pro plan at twenty dollars a month and runs on macOS or Windows only. Open it, click Cowork at the top, then make two choices in the model picker.
Select Opus 4.6 and turn extended thinking on. The picker describes Opus 4.6 as most capable for ambitious work, Sonnet 4.6 as most efficient for everyday tasks and Haiku 4.5 as fastest for quick answers, with extended thinking offered separately as a way to think longer on complex tasks. AI Central's instruction is to take the most capable option and the slower one, which is right when the job involves reading your files and planning before writing anything.
Build one folder, then stop
Step three is the spine of AI Central's whole method. Create one folder called Claude-Work, and put exactly four subfolders inside it. ABOUT ME holds who you are and how you write. PROJECTS gets one subfolder per live project. TEMPLATES holds your best past work, kept as structures rather than as finished pieces. OUTPUTS is where Claude saves finished files.
AI Central gives one reason for the discipline, and it is a safety point rather than a tidiness one.
Cowork has read/write access.
The follow-on instruction is two words, keep it tight. A tool that can write into a folder should be pointed at a small, deliberate folder. Four subfolders is a scope decision wearing the costume of a filing system, and it is the difference between a workspace you can audit and a drive you have quietly handed over.
Three context files beat fifty uploads
Inside ABOUT ME, AI Central has you create three markdown files. One called about-me, covering what you work on day to day. One called my-voice, holding your tone plus two or three real writing samples. One called my-rules, where the operating constraints live, ask before starting, show a plan, never delete.
One great file beats 50 random uploads
That line from AI Central is the argument against the habit everyone brings over from chat tools, which is dumping every document you own into context and hoping something sticks. Two or three genuine writing samples outperform a folder of PDFs because they are examples of the target output, not background noise.
The rules file earns its place for a different reason. It is the only spot in the setup where a permanent no lives, and a rule written once in a file is enforced far more consistently than one you remember to repeat when you are in a hurry.
Standing instructions, written once
Step five is where AI Central's setup stops being folders and starts being behaviour. You open the Cowork instructions editor in settings and paste a short standing brief that names who you are and your role, tells Claude to read your files before every task, to ask questions before executing, to show a plan, and never to delete without approval.
Set once, runs forever
Those four clauses do four separate jobs, which is why AI Central bundles them rather than picking one. Reading files first is what makes the folder worth building at all. Asking before executing is what stops a confident wrong answer. Showing a plan gives you a cheap place to intervene, before the work exists rather than after. Requiring approval to delete is the one that matters most in a folder the tool can write to.
The prompt that makes Claude interrogate you
Step six is AI Central's per-task wrapper, and it is the sharpest idea in the whole sequence. Every task opens the same way. Name the task, tell Claude to read all files first, and require it to ask you questions using AskUserQuestion before it executes, generating new questions if something is off, and never guessing.
What comes back is not a paragraph of polite clarification. Claude generates clickable buttons, you click, it executes. The interaction inverts, and AI Central states the payoff in five words.
Claude forces you to be clear
This is the part worth taking even if you never install Cowork. Vague instruction is the largest single cause of disappointing AI output, and the standard advice is to tell people to write better prompts. AI Central's version moves that burden onto the tool instead. You do not have to arrive with a precise brief, you have to answer the questions that produce one, and answering is far easier than authoring. Because the questions arrive as options you click, specifying a task becomes something you do every time rather than only when it feels worth the typing.
Plugins and connectors, deliberately last
The final two steps extend Cowork outward, and AI Central holds them until the foundation exists. Plugins come first, browsed and installed from the customize menu. Once one is installed, typing a forward slash surfaces the commands it adds, so a marketing plugin contributes a draft-content command, a data plugin an explore command, while a legal plugin can respond to a plain instruction like asking it to review an NDA.
AI Central's recommendation is to start with two. Connectors come last, added from the connectors section of settings, and the catalogue runs past fifty, with Notion, Gmail, Google Calendar, Slack, Figma and Canva among the most popular. Claude's connector page notes they are built by third parties and reviewed by Anthropic for safety, and that a custom one can be added too.
Restraint here is the actual advice. Every plugin adds commands and every connector adds reach, and both are far more useful pointed at a tidy folder and a written set of rules than sprayed across a machine that has neither.
What to do with this
How you should use AI Central's eight steps depends on where you are starting from.
- New to Cowork, do steps one through five in one sitting. The app, the model, the folder, the three context files and the standing instructions. Nothing after that works properly without them.
- Already using Claude daily, skip ahead and wrap one real task in the per-task prompt this week. Watching it ask you questions before it starts will teach you more than reading another set of instructions.
- Running a team, standardise the folder structure and the rules file rather than the prompts. Structure transfers between people, clever wording does not.
- Whatever your level, add plugins two at a time and add a connector only when a specific task needs one. AI Central's ordering is a warning about scope, not a limit on the tool.
Do I need a paid Claude plan to use Cowork?
Yes. AI Central's setup calls for a Pro plan at twenty dollars a month, and Cowork runs as a desktop app on macOS or Windows only, so there is no browser-only route and no phone route.
Which model should I pick in Cowork?
AI Central says Opus 4.6, with extended thinking switched on. The alternatives in the picker are Sonnet 4.6, described as most efficient for everyday tasks, and Haiku 4.5, described as fastest for quick answers. Both favour speed, which is the wrong trade when you are asking Claude to read a set of files and plan before acting.
What exactly goes in the Claude-Work folder?
Four subfolders and nothing else at the top level. One for who you are and how you write, one holding a subfolder per live project, one for templates drawn from your best past work, and one where Claude saves finished files. AI Central's instruction is to keep it tight, because Cowork has read and write access to whatever you point it at.
Why should Claude ask me questions before it starts?
Because a wrong guess is expensive and a question is cheap. AI Central's per-task prompt makes Claude read your files, then ask clarifying questions as clickable options before executing, and generate new ones if something looks off. The framing is that you do not need to arrive clear, the process makes you clear.
How many plugins and connectors should I install at the start?
Two plugins, according to AI Central, and connectors only as real work demands them. The connector catalogue runs past fifty, which is precisely why installing broadly on day one is the wrong first move.