Community Q&A

Cowork mode is a feature inside the Claude desktop app that lets Claude actually do things on your computer — not just chat. Instead of just answering questions, Claude can open and edit your Word documents, create Excel spreadsheets, build PowerPoint presentations, search the web, organise your files, and a lot more. Think of it like hiring a very capable assistant who sits at your desk and does the work for you. You describe what you want, Claude does it. No coding, no technical knowledge needed.

They're different products from different companies. ChatGPT is made by OpenAI, Claude is made by Anthropic. Both are AI assistants you can chat with — but Cowork mode specifically is Claude's ability to take actions on your desktop: editing real files, running code, connecting to apps like Slack or Google Drive. ChatGPT has similar agentic features but they work differently. If you're deciding between them, the best thing is to try both — but Cowork mode is particularly strong for document work and file management.

The Claude desktop app itself is free to download. However, Cowork mode (the feature that lets Claude take actions on your computer) requires a Claude Pro subscription or higher. As of early 2026, Pro costs around $20/month. Anthropic occasionally changes their plans so always check the current pricing at claude.ai. There's no free tier for Cowork mode itself, but the standard Claude chat (without desktop actions) is available on the free plan.

Getting started is straightforward: 1) Go to claude.ai and create a free account. 2) Download the Claude desktop app for Mac or Windows from claude.ai/download. 3) Open the app and sign in. 4) Look for the Cowork or "Computer use" option in the interface — it may be labelled slightly differently depending on the version. 5) Select a folder on your computer that you want Claude to work with. That's your workspace. Check out our Getting Started guide for a full walkthrough with screenshots.

Anthropic updates Claude regularly, so the interface does change over time. If the layout looks different from screenshots you've seen, it's likely because the screenshots are from an older version, or you're on a different plan. The core idea is always the same though — there will be a way to start a new conversation and a way to enable the "agentic" or "Cowork" mode. If you genuinely can't find the Cowork feature, check that you have a Pro subscription, and that you're using the desktop app (not just the browser version at claude.ai).

The browser version (claude.ai) is for chatting — you can ask questions, write content, analyse text you paste in. The desktop app is where Cowork mode lives. It has extra powers: it can access files on your actual computer, run code in a secure sandbox, connect to external apps via MCP, and take multi-step actions autonomously. Think of the browser version as a smart chat window, and the desktop app as a capable digital assistant who can actually open your files and do work.

This is a great question to ask. Cowork mode runs inside a secure sandboxed environment — a lightweight virtual machine that's isolated from the rest of your system. You control which folder Claude has access to, and it can only work within that folder unless you explicitly give it more access. Claude can't silently send your files anywhere or access things outside its workspace. That said, use common sense: don't give it access to folders with passwords, tax documents, or sensitive personal data unless you need to. Start with a test folder to get comfortable.

A lot more than most people expect. Common tasks include: creating and editing Word documents, Excel spreadsheets and PowerPoint slides; organising and renaming files; searching the web and summarising what it finds; writing and running Python or JavaScript code; sending emails or Slack messages (with the right connectors); generating reports from data; scheduling recurring tasks; and even building custom workflows that chain multiple steps together. If you can describe it clearly, there's a good chance Claude can do it.

Macros and Power Automate are rule-based — you define exactly what happens step by step, and they follow those rules. They're great for repetitive, predictable tasks. Cowork mode is intelligence-based — you describe the outcome you want and Claude figures out how to get there. It can handle ambiguity, make judgement calls, deal with unexpected file formats, and adapt when something doesn't go as planned. The tradeoff: macros are more reliable for simple repeatable tasks; Claude is better for complex, messy, or one-off work that's hard to script.

In Cowork mode, you don't drag files into the chat window — instead, you select a folder when starting your session. Look for a "Select Folder" or folder picker option in the Cowork interface. Once you choose a folder, Claude has access to everything inside it. Then you just tell Claude what you want: "read the file called report.docx and summarise it" or "edit the spreadsheet called sales.xlsx". If you want to work on a specific file, just move it into your selected folder first.

Yes — Claude can open, read, and edit Excel (.xlsx) files. It can add data, restructure columns, create new sheets, and write formulas including complex ones like VLOOKUP, SUMIF, INDEX/MATCH, and pivot logic. It uses Python libraries under the hood (like openpyxl) to manipulate the files. One thing to know: it works best when you describe what you want the outcome to be rather than the exact steps. For example: "Add a column that calculates profit margin from columns C and D" works better than specifying exact cell references.

Yes, Claude can read PDFs. It can extract text, summarise content, pull out specific data, and answer questions about what's inside the document. For PDFs with tables, it can usually extract those too. One limitation: if the PDF is a scanned image (rather than text-based), extraction is harder and less reliable. For editing PDFs (adding content, changing text), Claude can do this too but it works better when converting to a different format first, making the changes, then converting back.

Absolutely — this is one of Cowork mode's strengths. Claude can write a Python script to rename, move, sort, or reorganise hundreds of files based on any pattern you describe. Tell it the pattern (e.g. "rename all files to include today's date as a prefix, like 2026-02-19_filename.pdf") and it will write and run the script in your workspace. It will typically show you a preview before making changes. For large batches, it's good practice to say "show me the first 10 renames before you do all of them" so you can verify the logic is right.

Yes. Claude can merge multiple .docx files into a single document, optionally adding page breaks between each, preserving formatting, adding a table of contents, and even standardising styles across all files. Just put all 15 files in your workspace folder and say something like "Merge all the .docx files in this folder into one document called Combined_Reports.docx, add a page break between each, and create a table of contents at the start." It handles it using Python's python-docx library.

The cleanest approach is via MCP connectors. If you have the Google Drive or SharePoint MCP connector set up in Claude, it can browse, read, and write files directly from cloud storage without you downloading anything first. If you don't have those connectors, the workaround is to sync your cloud folder locally (Google Drive for Desktop, OneDrive sync) and point Cowork mode at the local sync folder — changes Claude makes will sync back to the cloud automatically. For large-scale operations, the MCP route is more reliable and doesn't require local disk space.

Yes! Cowork mode supports scheduled tasks. You can set Claude up to run a task at a specific time, daily, weekly, or on a cron schedule. For example: "Every weekday at 8am, check my inbox summary and create a to-do list for the day." You set this up once, and Claude runs it automatically. You can manage scheduled tasks from within the Cowork interface. This is one of the most powerful features for people who want to automate recurring work without writing any code.

This happens sometimes, especially with very long tasks. A few things to try: 1) Just say "continue" or "keep going" — Claude will often pick up where it left off. 2) Ask "what have you done so far and what's left?" to get a status update. 3) If it seems confused, ask it to summarise the current state and then give it a specific next instruction. For very long tasks, it's better to break them into smaller chunks upfront: "Do steps 1-3, then stop and tell me what you've done before continuing."

The best way is to create a template file in your workspace and reference it in your instructions. For example: "Create this report using the same format as template.docx in the workspace folder." Even better, create a SKILL file — a simple text file that describes your standard format, tone, and structure. Then tell Claude "follow the instructions in my SKILL file for all documents." This way you set the standard once and it applies every time. Check out the Skills guide for how to set this up.

Yes, with the right MCP connectors. If you have the Gmail, Outlook, or Slack MCP connected to your Cowork session, Claude can send messages as part of a workflow. For example: "Process the sales report, create a summary, then email it to team@company.com." You'll be asked to confirm before messages are sent (for safety). Without connectors, Claude can draft the email and open it in your email client, but you'd need to hit send yourself. Setting up connectors takes a few minutes and unlocks a lot of automation possibilities.

For reliable pipelines, the key is structured checkpoints and error handling in your instructions. Structure your prompt as: "Complete step 1 [description]. When done, output a JSON summary of what you did and any issues. Then proceed to step 2 only if step 1 succeeded. If it failed, stop and report the error." This forces Claude to produce verifiable intermediate outputs. For production-grade pipelines, combine this with scheduled tasks and skill files that define expected inputs and outputs at each stage. The more explicit you are about success criteria, the more reliably Claude chains steps.

Yes. Claude can write and execute Python code that calls any REST API — POST to a webhook, GET data from an API, authenticate with OAuth tokens, etc. You'd typically say: "Write a Python script that calls this API endpoint [URL] with these parameters and saves the response to results.json." Claude handles the code, runs it in the sandbox, and shows you the output. For recurring API calls, combine this with scheduled tasks. For sensitive API keys, store them in a local config file in your workspace and tell Claude to read the key from there rather than pasting it in the chat.

A Skill is like a set of instructions Claude reads before starting a task. It's a file (usually called SKILL.md) that tells Claude things like: how you like your documents formatted, what your brand voice sounds like, what steps to follow for a specific process, or what context it needs about your business. Once you set up a Skill, Claude follows those instructions automatically every time that Skill is active — you don't have to repeat yourself. It's the difference between briefing an assistant from scratch each day versus having a trained staff member who already knows your preferences.

Good question — they're related but different. A Skill is a set of instructions (a text file) that guides how Claude behaves for a specific type of task. A Plugin is a packaged bundle that can include Skills, MCP connectors (connections to external apps), and other configuration — all bundled together and installable in one click. Think of it this way: a Skill is like a recipe, and a Plugin is like a meal kit that includes the recipe plus all the ingredients. Plugins are great for sharing complete workflows with a team.

A good SKILL.md typically includes: 1) Purpose — what this skill is for. 2) Context — background about your business, role, or project that Claude needs to know. 3) Instructions — step-by-step guidance on how to perform the task. 4) Format rules — how outputs should look (tone, structure, length). 5) Examples — sample inputs and outputs if possible. Write it in plain English, not code. The clearer and more specific you are, the better Claude follows it. Start with 1-2 pages and refine it over time as you notice Claude missing things.

Yes — Skills are just files, so you can share them like any document. The easiest ways: put the SKILL.md in a shared folder (Google Drive, Dropbox, SharePoint) and have teammates copy it to their Cowork workspace; or package it as a .skill file (a zip archive) and share it directly. If you're on Claude Teams, you may also be able to share Skills through the team's plugin/skill library. The key is making sure the instructions in the Skill are general enough to work for different people, not hard-coded with your personal details.

For complex workflows, structure your SKILL.md with: a Metadata header (name, version, description, trigger conditions); Preconditions (what inputs must exist before starting); Numbered phases with clear success criteria for each; Error handling rules (what to do when something fails); Output specification (exact format of what to produce); and a Verification step (how Claude should check its own output). Use explicit language like "You MUST" and "You MUST NOT" for non-negotiable rules. Versioning your SKILL.md lets you roll back when a change makes things worse.

The biggest tip: describe the result you want, not the steps to get there. Instead of "click on this, then do that", say "I want a two-page sales summary document with bullet points for each product, sorted by revenue." Also: give context. Claude doesn't know your business or your goals — a sentence of background makes a huge difference. If it gets it wrong, don't start over — just say "that's not quite right, I wanted X instead of Y." Claude learns from corrections within the conversation. Shorter, clearer prompts usually beat long complicated ones.

Give Claude an example of the output you want. If you want a report formatted a certain way, show it a previous report you liked. If you want emails in a specific tone, paste in an example. This is faster and more reliable than trying to describe the format in words. Claude is very good at pattern-matching — if it can see what "good" looks like for you, it can replicate it. This one habit will improve your results more than anything else, for any type of task.

Just tell it directly. Start your prompt with: "Just do the task, don't explain what you're doing." Or: "Skip the commentary, just produce the output." Claude responds well to explicit tone and style instructions. You can also put this in a Skill file so it applies every session: "Always produce output directly without explaining your reasoning unless I specifically ask for it." Once you set this expectation, Claude usually sticks to it throughout the conversation.

Yes — add a verification instruction at the end of your prompt. Something like: "Before you finish, re-read the output and check: (1) does it match what I asked for, (2) are there any obvious errors, (3) is the formatting correct? Fix anything you spot, then give me the final version." This simple addition catches a surprising number of mistakes. For critical work, go further: "Verify the numbers by recalculating them" or "Check that every section of the document has been completed." Claude is better at catching its own errors when you explicitly ask it to look.

A few good approaches: 1) Skill files — put your common instructions in a SKILL.md so they're automatically active each session. 2) Prompt library — keep a simple text file called prompts.txt in your workspace with your go-to prompts. Tell Claude "use prompt #3 from my prompts.txt file." 3) Scheduled tasks — for prompts you run on a schedule, set them up as recurring tasks. The Skill approach is the most powerful because Claude actively reads and follows those instructions rather than you having to copy-paste each time.

A few that don't get enough attention: 1) Sub-agents — you can have Claude spawn specialist agents for different parts of a task and run them in parallel. 2) Session transcript reading — Claude can read transcripts from previous sessions to maintain continuity across days. 3) Worktree isolation — for code changes, running in an isolated worktree means experiments can't break your main files. 4) Skill chaining — having multiple Skills active at once, each handling a different domain. 5) The verification step pattern — ending every task instruction with a self-audit prompt dramatically reduces errors on complex outputs.

A few common causes: 1) The folder might be in a location the sandbox can't reach (like an external drive or a network location) — try moving your files to a local folder like Desktop or Documents. 2) You may need to restart the Cowork session after selecting the folder. 3) Check that the file names don't have unusual characters or are extremely long paths — this can cause access issues. 4) On Mac, check System Preferences → Privacy & Security → Files and Folders to ensure the Claude app has permission to access your folder.

A few possibilities: 1) Phrasing — sometimes rephrasing helps. Instead of "send this email" try "draft this email and send it using Gmail." 2) Missing connector — if you're asking it to interact with an app (Gmail, Slack, etc.), you may need to connect that app via an MCP connector first. 3) Safety guardrails — Claude won't do certain things like enter passwords, delete files permanently, or share sensitive data. These are intentional safety limits. 4) Context loss — in a long session, Claude can sometimes forget earlier permissions. Try restating the task clearly in a new message.

This usually happens because Claude built the document using python-docx which doesn't have access to all fonts. Try these fixes: 1) Tell Claude "use standard fonts only — Arial, Calibri, or Times New Roman." 2) Give Claude a template file: "Use template.docx as the base for formatting" — it will inherit that file's styles. 3) If the document structure is broken, ask Claude to rebuild it: "The document has formatting issues, please recreate it cleanly from scratch using only default Word styles." Templates are the most reliable long-term fix.

The direct approach: paste the error message into the chat and say "Fix this error and try again." Claude is good at debugging its own code when given the actual error. If it's gone in circles, try: "Start fresh. Forget the previous approach. Here's what I want the script to do: [describe the goal]. Write a new clean version." The sandbox is isolated so a failed script can't damage anything — you're just losing time, not data. For future prevention, ask Claude to add error handling and print statements to long scripts so failures are easier to diagnose.

Scheduled tasks run in a fresh session each time, so you need to make sure your task instructions explicitly specify the save location. If the instruction just says "create a report," Claude might create it but not know where to save it in a fresh session. Fix your scheduled task prompt to say: "Save the output to [specific folder path] as [filename with date]." Also check the task's session transcript — you can review what Claude actually did by reading the session log. If the file was created but not in the right place, it's a path specification issue in your instructions.

Inconsistency usually comes from one of three causes: 1) Ambiguous instructions — Claude is filling gaps differently each time. Go through your SKILL.md and replace any vague language ("make it professional") with specific rules ("use formal English, no contractions, sentences under 25 words"). 2) Missing context — the Skill assumes Claude knows something it doesn't. Add more explicit background. 3) Conflicting instructions — two rules contradict each other and Claude resolves them differently. Add priority ordering: "Rule A takes precedence over Rule B." Run the same test prompt 3-4 times and compare outputs to isolate which instruction is causing variance.