Claude Cowork: The Complete Guide (2026)

For most of 2025, using AI meant getting really, really good at copying and pasting.

You'd copy your data into ChatGPT or Claude. You'd copy the answer back out into Excel, Word, or an email. I used to call this the first-mile problem (getting your data in) and the last-mile problem (getting your data out). It was magical and maddening at the same time.

Claude Cowork kills the copy-paste. Instead of you ferrying files back and forth to a chatbot in the cloud, Cowork sits on top of your actual files — on your actual computer — and does the work. It reads your folders, writes Word docs, builds spreadsheets, makes slides, browses the web, and runs tasks while you sleep. No uploads. No downloads. No copy-paste.

It's also genuinely hard to learn. There are a lot of moving parts — memory, connectors, projects, artifacts, skills, scheduled tasks, the Chrome extension, the Excel and PowerPoint add-ins — and Anthropic ships updates multiple times a day. This guide is my attempt to put the whole thing in one place, in plain English, for the knowledge worker who has never written a line of code.

A quick note on where this comes from: I teach Claude Cowork to hedge funds, private equity firms, and asset managers, and I built an entire course on it. The questions and framings below are pulled straight from those live trainings — so this is the stuff people actually get stuck on, not a feature list.

The big idea: The frontier models are now roughly equally smart. The differentiation isn't intelligence anymore — it's the product surface: how that intelligence plugs into your everyday work. Cowork is Anthropic's bet on owning knowledge work the way Claude Code owns software engineering.

What is Claude Cowork?

Claude Cowork is a feature inside the Claude desktop app — not a separate product you buy, and not something you can use in a browser tab. When you open the desktop app you'll see three tabs: Chat, Cowork, and Code. Code is the coding agent for software engineers. Chat is the conversational interface you already know. Cowork is the new one, and it's aimed squarely at people who don't code but still want access to deep, agentic workflows.

The reason it's desktop-only is the whole point of the product: Cowork works directly with the files on your machine. It needs to be able to open Finder folders, read your documents, and save outputs back to disk. A web tab can't do that.

The shift Cowork represents is from "turn-and-burn" answers — you ask, it answers, you ask again — to long-running agentic tasks. Anthropic's most capable Opus models can now work autonomously for a long stretch on a complex task without you holding its hand the entire way. (The research lab METR tracks how long AI can work unsupervised, and the curve has gone vertical with the latest frontier models.) Cowork is where that capability meets your real work.

Cowork vs. Chat vs. Code

Here's the cleanest way to hold the distinction in your head:

	Chat	Cowork	Code
Where your files live	The cloud (you upload)	Your computer	Your computer / repo
Can it edit your files?	No — it outputs, you download	Yes — read, create, edit, rename, delete	Yes
Who it's for	Everyone	Knowledge workers	Software engineers

The headline difference: Chat plays in the cloud; Cowork touches your files. In Chat, you upload a document, get an output, and download it — the file never really lives anywhere. In Cowork, you point Claude at a folder and it can read everything in there, rename files, reorganize them, create new files, and edit existing ones in place. You never upload anything.

I suspect Chat and Cowork will eventually merge. But today they're separate tabs, and knowing which one you're in matters.

Is Cowork safe? Can it delete my files?

This is the single most common fear in every training I run, so let's hit it head-on.

Cowork operates inside a walled garden. When you start a session, you pick exactly one folder for Claude to work in. From that point:

• It can only act inside that folder.
• It can go deeper — create subfolders, organize things into them.
• It cannot go up the hierarchy to your wider drive. The parent folders are invisible to it.

So the blast radius is whatever folder you chose, and nothing else. There are also guardrails around deletion specifically, and you control a permission system (more on that in the Connectors section) that lets you require approval before any action.

My #1 tip for nervous beginners: Start with a sandbox copy. Duplicate the folder you want to work on, point Cowork at the copy, and go to town. Once you trust what it does, work on the real thing. This removes 100% of the "what if it breaks something" anxiety.

Getting set up

1. Download the desktop app. Cowork lives in the Claude desktop app for Mac and Windows. (Reminder: not the web version.)
2. Keep it updated. Anthropic ships fixes constantly. Check the top-left corner of the app for updates regularly — a lot of "bugs" are just old versions.
3. Get on a paid plan. Cowork is token-hungry. The entry-level Claude plan works for light use, but you'll burn through it fast once you're doing real document and presentation work. The next tier up gives you a lot more headroom. Teams and Enterprise plans set usage at the administrator level.
4. Pick your model. When you're learning, use the most capable Opus model with extended thinking turned on. You want to see the full potential first. Later, you'll downshift to the faster, cheaper Sonnet for daily work and save Opus for genuinely hard problems. (More on this in the cost section.)

How memory works

Here's a truth that trips everyone up: LLMs don't remember anything. Every new conversation starts from a blank slate. They are, by default, stateless.

The classic analogy is that an LLM is like an intern — but that's where it breaks down. If you told a human intern "use 14-point Arial, not 12-point Times," they'd remember next time. An LLM won't. So Claude does something clever in the background: it reads across your past chats and auto-generates a memory file — its best guess at what it should know about you (your work context, your role, your preferences).

You can see and edit this. Go to Settings → Capabilities:

• Turn on Search and reference chats so Claude can pull in past conversations on demand.
• Turn on Generate memory from chat history to get the auto-generated profile.
• Use the pencil icon to tell Claude to remember or forget specific facts ("I have three kids," "forget that this was a client").

Two things to know. First, memory is lossy — it's inferring, not transcribing, so it'll miss things unless you explicitly say "remember this." Second, it's partly osmosis: you can't perfectly audit how it's being applied. When in doubt, just put the important fact in your prompt or your instructions.

Global and custom instructions

Cowork has a dedicated Global Instructions file (Settings → Cowork → Global Instructions). Every time you run a prompt in Cowork, this text gets injected into the prompt automatically. Think of it as a standing override of how you want things done.

Good things to put there:

• Your role ("I'm an investment analyst at a long/short fund").
• Output preferences ("For text, use Markdown. For spreadsheets, use Excel. For decks, use PowerPoint.").
• Firm-specific jargon or acronyms.
• Safety rules ("Never delete a file without checking with me first.").

Use this real estate carefully. It's appended to every single prompt, so keep it general and lean. If you stuff it full of detail, it starts to override your actual requests. Don't overthink it — but if you catch yourself typing "save it as Excel" for the hundredth time, that belongs here.

Projects have their own custom instructions that work the same way, scoped to that project (covered below).

Connectors (the MCP advantage)

Connectors are how Claude reaches outside your files into the apps you use every day — Gmail, Google Drive, Slack, Notion, your calendar, your CRM, and industry tools like S&P, FactSet, and PitchBook. This is what turns Claude from a file tool into something closer to an operating system for your work.

Under the hood they run on MCP (Model Context Protocol) — you don't need to know the acronym, but it's the open standard Anthropic pioneered that translates your plain-English request ("read my email from John") into the right API call, handling all the authentication and permissions. You only ever get access to what you already have access to.

Two things make connectors powerful:

1. They're not read-only. You can pull a folder of research in and create a Gmail draft, a Notion page, or a CRM entry. Read, write, and act.
2. You can combine them in one request. "Pull my last five Granola meetings and add the key points to a new Notion database" hits two connectors in a single natural-language prompt.

The permission model

This is the safety system, and it's worth understanding. Every connector is a bundle of tools — and there are always more reading tools than writing tools (it's safer that way). Gmail, for example, has several read tools (search, read thread) but only one write tool: create draft. Claude literally cannot send an email — the most it can do is draft one.

For each tool, you (or your admin) set one of three permission levels:

• Always allow — runs without asking.
• Ask for permission — prompts you each time.
• Blocked — can't run at all.

My approach: Start by blocking all the write tools, so Claude can read but not create. For the read tools, set "ask for permission." Then, as you get comfortable, loosen to "always allow." Safe by default, faster over time.

Projects

If 2025 was all about prompts, this year is all about context. The models can now ingest far more of it — but for Cowork users, that context is scattered across countless files and folders. Projects are how you wrangle it.

A Project is a self-contained workspace — think of it as a folder — that bundles four things:

• Files and folders that serve as persistent context for every chat in the project.
• Custom instructions that apply to every task (tone, format, what "done" looks like).
• Background memory that refreshes periodically from your work.
• A natural home for scheduled tasks.

The standout idea is scoped memory: what Claude learns in one project doesn't leak into another. That's both a productivity feature and a privacy guardrail. As Anthropic puts it: "Projects keep your tasks and context in one place... Files and instructions stay on your computer."

A useful mental model: what LLMs actually remember

Not all context is equally reliable. Roughly, from most to least reliable:

1. Most reliable (like RAM): your prompt, your follow-up messages, and files you just uploaded. Always there — but limited in size.
2. Middle ground: project files. Robust and persistent, though not effortless to update.
3. Least reliable: long-term auto-memory. Helpful, but osmotic — you can't fully curate it.

Projects sit in that sweet spot between "what's in your prompt" and "what we hope the model remembers."

Setting one up

In Cowork, go to the Projects tab and you'll get three creation paths:

• Import an existing Claude (chat) project — pulls in its files and instructions.
• Start from scratch — name it, set a save location, add instructions and files.
• Use an existing folder — point Cowork at a folder you already have and layer instructions on top.

One subtle but powerful detail: the folder that holds your context can be different from the folder where outputs get saved. That's a big deal for teams — everyone reads from a shared context folder but writes outputs somewhere else.

Best practice: name projects by outcome, not topic. "Tesla" is too broad — you'll blow past the file limit and drown in noise. "Monthly Reporting — Tesla" is scoped to a deliverable. Start narrow, expand once you find the right amount of context.

Artifacts

Artifacts are interactive, shareable mini-apps that Claude builds for you. Technically each one is a single file of code that runs in a browser, so you can publish it to a URL and anyone can use it — no server, no setup. Build once, share infinitely.

Three examples I love from my own use:

• An interactive employee handbook. Upload the PDF, ask Claude to turn it into a shareable chatbot that answers policy questions (and says "I don't know" when the answer isn't in the doc). Now HR has a self-serve FAQ.
• A data-visualization dashboard. Feed it a dataset (say, housing starts and the S&P 500) and ask for an interactive artifact with time-series charts, a correlation matrix, and normalized performance views. You can toggle and explore right in the browser.
• A game — with your kid. "Create an artifact that's the classic Tetris game." My daughter went from this to building and downloading her own games into a portfolio site. Games are one of the fastest on-ramps to AI for non-technical people of any age.

The deeper point: artifacts let non-technical colleagues use a specialized tool without knowing any AI themselves. You build it; they just click the link.

The Chrome extension

As we entered 2026, the battle for AI supremacy was everywhere. ChatGPT vs. Claude. Midjourney vs. Nano Banana. Nvidia's Blackwell chips vs. Google's TPUs. Codex vs. Claude Code. And then there's the battle for the soul of your computer: the web browser.

Claude's Chrome extension lets Claude take control of your actual browser — clicking buttons, filling forms, navigating pages, extracting data, and running multi-step workflows. Because it's tied to your account, it inherits your memories and your skills. And it's ready with a single click.

It has two modes, and the distinction matters:

• Ask before acting — it shows you its plan and waits for approval at each step.
• Act without asking — it just goes.

Start with "ask before acting." This thing controls your browser, so you want to watch it before you trust it.

Four tasks that show the range:

1. Booking a haircut on Calendly. "Book me a spot the week of the 20th in the afternoon." Done in two minutes.
2. Researching flights. "Direct flights from LAX, premium economy, somewhere wintery in February." It spent 24 minutes driving Google Flights, adjusting dates and cabins. Output was excellent — though one of my own manual searches caught an option it missed (Munich). Always verify.
3. A recurring scraper. "Look up BTC, ETH, and SOL prices every hour and put them in a Google Doc." It created a "Crypto Price Tracker" doc; the three-dots menu turns it into a recurring task.
4. Summarizing long articles. The most obvious use, and the one that quietly changed my reading habits — I now "read" far more long pieces because the friction is gone.

The risk to know about: prompt injection. When Claude reads an untrusted web page, malicious text on that page could try to hijack it. Newer models are increasingly trained to resist this, and "ask before acting" is your seatbelt. Note also that the extension is blocked from financial institutions by design.

Skills

Skills are the feature that, in my opinion, separates AI power users from everyone else. And the reassuring part: a skill is just a good old-fashioned text file.

A skill is a set of instructions — packaged as a simple folder — that teaches Claude how to handle a specific recurring task your way, every time. The core file is named exactly SKILL.md (case-sensitive) and has two parts:

• YAML frontmatter — tells Claude when to use the skill. The description field must say both what it does and when to use it, in under 1,024 characters. This is what lets Claude trigger the skill automatically.
• Markdown body — the actual instructions Claude follows when the skill runs.

Optionally, the folder (named in kebab-case) can include references/, scripts/, templates/, and assets/ subfolders.

Progressive disclosure (why skills are efficient)

Skills don't dump everything into context at once. They load in three levels:

1. Frontmatter — always loaded, just enough for Claude to know a skill exists and when it's relevant.
2. The SKILL.md body — loaded only when Claude decides to use the skill.
3. Linked files — opened only as needed.

This keeps your context window lean while giving Claude deep capability on demand.

How to make one

Use the skill-creator skill — Anthropic's meta-skill for building skills. You don't touch code or Markdown directly. You just describe your workflow in plain English ("create a skill that does X"), and it writes the SKILL.md for you. Then you test it, refine it ("make the jokes edgier," "add a summary table at the top"), and save it to your library.

Watch for this: When you test a skill, confirm Claude says it's reading the skill before it answers. If it doesn't, it's falling back to generic behavior and you've lost the benefit.

Worked examples from my own library

• Research analyzer — drops a 50-page broker report into a consistent format: price targets, key takeaways, company-by-company sentiment, downside scenarios.
• Style guide — replicates a specific writing voice (I trained one on Howard Marks's memos) so every newsletter or memo sounds right.
• Inbox triage — color-codes your email red/yellow/green by what actually needs a response.
• Transcript parser — turns a Zoom or Granola transcript into action items, objections, pain points, and a deal-close likelihood.

And they compose: my transcript parser feeds my proposal-generator skill, turning a sales call into a draft proposal email. What used to take 30 minutes takes one.

Skills vs. Projects

Simple split: Projects manage context (the folders and files). Skills codify workflows (the repeatable how-to). Use them together — a project holds the material, a skill runs the same process on it every time.

Scheduled tasks

For all of 2025 I argued that chat wouldn't be the main AI interface. The one missing piece was automation — Claude doing work on a schedule, unattended. Scheduled tasks close that gap. They're essentially CRON jobs you can click — no code required.

Two ways to create one:

• Type /schedule in a task's chat input and let the interactive skill walk you through it.
• Go to the Scheduled page in the sidebar → + New task, and fill in the name, prompt, frequency (hourly, daily, weekdays, weekly, or on-demand), model, and working folder.

A scheduled task can use your files, skills, and connectors together. Two of my favorites:

• Meeting prep. Every morning, look at tomorrow's calendar; for each external meeting, pull a short company description, the person's recent LinkedIn posts, any past Granola transcripts, and relevant email history; push it all to a Slack channel.
• Twitter/X scan. Each day, collect the top AI tweets by engagement and append them (no duplicates) to a Notion database.

The gotcha everyone hits: Scheduled tasks run locally, which means your computer must be on and awake with the desktop app open. If your laptop is closed in a bag, the task won't fire — it'll wait until you're back. There's a setting to keep your machine awake during runs, but travel still breaks it.

A practical tip: use "Run now" to test a task once. The first run is when Claude asks for all its permissions, so babysit it — after that, scheduled runs are smooth. And schedule heavy tasks for the middle of the night, when token usage is cheaper.

Excel and PowerPoint

The Excel and PowerPoint add-ins aren't just "Claude in a sidebar." They use a custom harness — a communication layer tuned specifically for Office — which makes Claude meaningfully better at spreadsheet and slide work. You install them from within Office; once open, Claude can read from connected files (an Excel model can feed a PowerPoint, and vice versa).

What it does well:

• Excel: build formula-driven models (I've had it build a clean five-period DCF with a sensitivity table, and a startup distribution waterfall showing dilution across funding rounds), extract data from messy PDFs into structured tables, and run analyses. Formulas are real, not hard-coded, so changing an assumption cascades correctly.
• PowerPoint: turn a workbook plus some source PDFs into a 7-slide deck — performance summary, winners and losers, key takeaways.

Set expectations on design. Output is roughly 80% there — clean and accurate enough for internal use, but if you work somewhere with a strong design culture, the polish won't match a human's. Excel is more reliable than PowerPoint here. And both add-ins are token-intensive (lots of formatting iterations), so watch your usage.

Dispatch and Computer Use

Two newer, more experimental capabilities:

• Dispatch lets you talk to your Cowork sessions from your phone — "what tweets did you pull today?" — while your desktop does the work. It's genuinely wild (you're talking to your computer from across the country) but currently slow and a bit unstructured. Early days.
• Computer Use lets Claude control native desktop apps — Finder, VLC, legacy software with no API — by taking screenshots and clicking around. Extremely powerful, correspondingly risky (you're handing over your whole machine), and currently banned on most corporate devices. Don't run it unsupervised.

Both are previews of where things are going more than daily drivers today.

What it actually costs (and how to spend less)

Cost is the most-discussed topic in every training I run, so let's be honest about it.

Consumer plans are heavily subsidized right now — think of the early Uber days when a ride to the airport was suspiciously cheap. A consumer plan caps your spend at a fixed monthly number, but the tokens you actually consume can be worth multiples of that. The subsidy won't last forever. Teams and Enterprise plans, by contrast, are pay-per-use from day one, so heavy Opus usage adds up fast.

The single most practical lever you have is file format. As a rough rule of thumb for how many tokens it costs Claude to process:

• Plain text ≈ 1
• Word doc ≈ 10
• PDF (especially with images) ≈ 100

If you can convert a stack of image-heavy PDFs to text first, you'll cut your costs dramatically. Other tips:

• Use Sonnet as your daily driver, Opus for hard problems. Don't drive the Ferrari to the supermarket.
• Keep chats short and focused. Long, sprawling context is expensive to re-process.
• Use Projects for persistent context instead of re-uploading.
• Batch heavy work overnight.

The reframe: in 2025 you learned to use these tools effectively. The next phase is using them efficiently — output relative to cost.

Frequently asked questions

These are the real questions people ask in my Claude trainings, lightly cleaned up.

Is Cowork a separate app or a feature inside Claude?

It's a feature inside the Claude desktop app — one of three tabs (Chat, Cowork, Code). It is desktop-only because it works directly with the files on your computer, which a browser tab can't do. If your screen looks nothing like a tutorial's, you're probably on the web version instead of the desktop app.

Can Cowork actually change or delete my real files? Is that safe?

Yes, Cowork can edit, rename, create, and (with guardrails) delete files — that's the whole point. But it can only do so inside the one folder you point it at. It can create subfolders (go deeper) but cannot climb up to your wider drive. For peace of mind, start with a sandbox copy of a folder until you trust it, and use the permission settings to require approval before actions.

Which folder should I point Cowork at?

Point it at the narrowest folder that contains what the task needs. A broad folder (your entire Dropbox) is slower, less accurate, and more expensive in tokens than a focused subfolder. Today a session works within one chosen folder; if you need several sources, a Project lets you assemble context more deliberately. When in doubt, duplicate the relevant folder as a sandbox first.

What's the difference between Cowork, a chat, a project, and a skill?

• Chat = conversational Q&A in the cloud; you upload and download.
• Cowork = the same intelligence working on your local files, able to create and edit them.
• Project = a workspace that manages context (files, instructions, memory) for a body of work.
• Skill = a reusable text file that codifies a workflow so Claude does a recurring task your way every time.

Rule of thumb: Projects manage context; skills manage process. Use a project to hold the material and a skill to run the same steps on it.

How much does Claude Cowork cost?

You need a paid plan. The entry plan works for light use but you'll hit limits quickly; the next tier gives much more room. Cowork uses more tokens than chat, so costs add up faster than you expect. The biggest savings come from using Sonnet as your default, avoiding image-heavy PDFs, and keeping chats focused (see the cost section above).

Can Cowork really rename and reorganize a folder of messy files?

Yes — this is the classic "first magic moment." Point it at a folder of badly named PDFs and say "find everything about each client and group it into folders," or "map this CSV into new columns matching my schema." Under the hood, Cowork is running Claude Code — it writes and executes small scripts to do the work — which is why it's so capable at file operations.

How do I make sure Cowork didn't make mistakes?

Build a verification habit. Ask it to check its own work before finalizing ("review the output against the source files and flag any discrepancies"), spot-check the result yourself, and bake a verification step into your global or project instructions so it happens every time. For anything important, treat Claude's output as a strong first draft, not gospel.

How do I keep my files organized so Claude works well?

Data hygiene is the unglamorous prerequisite. Decide on a small set of "source of truth" documents per project rather than dumping everything in. Use clear folder names, scope projects to outcomes, and prune files you no longer need. A messy folder produces messy results.

How do scheduled tasks work — do they run when my computer is off?

You set a prompt and a cadence, and Claude runs it automatically. But scheduled tasks run locally, so your computer must be on, awake, and running the desktop app. If it's off, the task is skipped and runs when conditions resume. A recurring task does not require you to change the date each time — that's the point of the schedule.

How do I create a skill, and how does Claude know when to use it?

Use the skill-creator skill and describe your workflow in plain English. Claude knows when to use a skill from its description (the "what and when" in the frontmatter) and triggers it automatically. A skill can contain multiple use cases, and you can tell Claude to use just one part of it. (A plugin is closer to a packaged extension that can include and share skills; a skill itself is just the instruction file.)

Can Claude run more than one task at once, and how do I track a long run?

You can kick off work and let it run while you do other things; long tasks show a live plan and to-do list on the side so you can see progress. If something's been spinning for many minutes, check that panel — it'll show what step it's on, which tells you whether it's working or stuck.

Will Claude send emails or take actions on my behalf automatically?

Not without you. The defaults are human-in-the-loop. Gmail, for instance, can only create drafts — it cannot hit send. For any write or action, the permission system lets you require approval. You decide how much autonomy to grant, tool by tool.

What can connectors do — can Claude read my email and calendar?

Yes. Connectors let Claude read and act across Gmail, Calendar, Drive, Slack, Notion, your CRM, and more — including reading email (and attachments) and proposing draft replies, or drafting a calendar invite that avoids Mondays and leaves a buffer. Whether a given connector is available depends on your plan and your administrator's settings.

How is the Chrome extension different from Cowork?

Cowork works on your files; the Chrome extension works on the web. The extension drives your browser (clicking, filling forms, extracting data) and can cross paywalls on sites you're already logged into. It typically asks permission the first time at a new site, then can act more freely. Because it's tied to your account, it shares your memories and skills.

How does memory work — what does Claude remember about me?

Claude auto-generates a memory profile from your chat history, but it's inferred and lossy — it often misses things unless you explicitly say "remember this." You can view and edit your memories in Settings → Capabilities, add facts going forward, and tell it to forget things. For anything critical, put it in your instructions rather than relying on memory.

Is my data safe, and where does it go?

This is partly plan- and policy-dependent, especially in regulated fields. Cowork keeps files and instructions on your computer; connectors only reach data you already have access to; and consumer vs. team/enterprise plans have different data-handling terms. If you're in a HIPAA- or compliance-sensitive role, confirm your organization's plan terms and your admin's connector policies before wiring up sensitive sources.

For teams and firms

A candid note for organizations: today, much of Cowork is effectively single-player. Skills, projects, and tasks live on an individual's machine, and sharing them across a team is still maturing (plugins are the emerging path for sharing skills). Administrators on Teams and Enterprise plans can control which connectors are available and set permissions firm-wide. If your goal is a shared "hive mind" of pooled context and standardized workflows across analysts, that's an organizational-design project as much as a tooling one — and exactly the kind of rollout I help firms with.

For everyone else, the bottom line is this: this is the worst these models will ever be. They'll only get faster, cheaper, and smarter. The people who push through the messy, jagged edge now — who build the skills, wire up the connectors, and develop real fluency — will have an enormous head start. Approach it with a creative and open mind, expect some frustration, and keep experimenting.

If you found this useful, my newsletter Future-Proof Your Career with AI goes deep on Claude, Cowork, and the rest of the AI toolkit every week. You can also browse my full AI Resource Guide.