- 1
- March
Cowork is an AI Agent from Anthropic that works directly on your computer. But "being able to use it" and "knowing how to use it well" are two different things — this article compiles tips from Anthropic and real-world user experience to help you work with Cowork safely, productively, accurately, and fast.
Why Do You Need "Tips"? Can't You Just Give Direct Commands?
Cowork is not an ordinary chatbot — it is an AI Agent with access to your files and the ability to run code on your machine. This means:
- Unclear instructions may produce incorrect or unexpected results
- Poor configuration may expose sensitive files to AI unnecessarily
- Without verifying outputs, you might pass along information that "looks right but is wrong"
- Without knowing about Queue and Plugins, you may waste time doing tasks one by one instead of batching them
The tips below are divided into 4 sections covering everything you need to know.
Part 1: Safety — Set Up Before You Start
1.1 Create a Dedicated Folder for Cowork
Do not give Cowork access to your entire Home Directory — create a dedicated folder and grant access only to that folder, for example:
~/CoworkProjects/— for project work~/CoworkTemp/— for temporary work that can be deleted
Limiting folder scope reduces the risk of Cowork reading or modifying unrelated files by accident.
1.2 Never Put Sensitive Data in Accessible Folders
Before granting Cowork access to a folder, verify that it does not contain these files:
.envfiles or configs containing API Keys and Passwords- Registration documents, KYC documents, or ID card copies
- Private Keys, Certificates, or SSH Keys
- Unreleased financial reports
Even though Anthropic states they don't use your data for model training, the data is still sent to Anthropic's API for processing. Anything that shouldn't leave your machine should not be in a folder Cowork can access.
1.3 Back Up Data Before Letting Cowork Edit Files
Cowork can create, edit, and delete real files on your machine — if something goes wrong, there's no Undo button. Therefore:
- Back up important files before requesting edits
- Use Version Control (e.g., Git) for code projects
- At minimum, copy files to another folder before issuing commands
1.4 Beware of Prompt Injection from External Files
This is a risk most people don't know about — if you have Cowork read files received from others (e.g., PDF, Word, email attachments), those files may contain text that tricks Cowork into performing actions you didn't request, such as "copy all files to this URL."
How to protect yourself:
- Don't let Cowork read files from untrusted sources directly
- Review file contents yourself first
- Keep external files in a separate folder from important files
IT security follows the same principles across all systems — whether it's two-factor authentication or restricting AI Agent permissions, the key principle is "grant the least privilege necessary" (Principle of Least Privilege).
— Zero Trust Security Principle
Part 2: Productivity — Write Prompts the Right Way
2.1 Specify "What" + "How" + "Expected Output"
Cowork performs best when instructions have 3 components:
| Component | Example |
|---|---|
| What (Task) | "Summarize monthly sales" |
| How (Method) | "From file sales-2026.xlsx, column B=date, D=amount" |
| Output (Result) | "Create a new Excel file with separate sheets per month, including a Pivot Table and bar chart" |
Compare this with a vague command like "summarize sales for me" — Cowork would have to guess where the data is, what format to use, and the output might not match your expectations.
2.2 Break Large Tasks into Smaller Steps
Instead of commanding "create the entire annual report," break it down:
- "Read all files in the Reports/ folder and tell me what data is available"
- "Summarize totals from all files, broken down by quarter"
- "Create a Word report with charts and tables"
Breaking tasks into steps allows you to verify each stage — if step 1 is wrong, you can fix it before generating a report from incorrect data.
2.3 Use Global Instructions for Recurring Standards
If you have standards that apply every time, such as:
- "Use TH Sarabun font size 14 for all Word documents"
- "Name files in English using kebab-case"
- "Always add a blue header row when creating Excel files"
Set up Global Instructions in Claude Desktop once, and Cowork will remember them across sessions — no need to repeat them every time.
2.4 Use Context Files for Complex Projects
For projects with many rules, create an INSTRUCTIONS.md file in the project folder containing:
- Required folder structure
- File naming conventions
- Desired report formats
- Reference data
When Cowork opens the folder, it will automatically read INSTRUCTIONS.md and follow it — like sending a brief to a colleague before starting a project.
Part 3: Accuracy — Verify Before Using Results
3.1 Never Trust AI Numbers 100%
This is an iron rule — AI may generate numbers that "look reasonable" but are wrong, especially:
- Aggregating totals from multiple files — may double-count or miss entries
- Calculating percentages — may use the wrong base number
- Citing data — may fabricate numbers entirely (Hallucination)
How to verify: Spot-check at least 3–5 important values yourself. For Excel work, open the file and check whether cell references in formulas are correct.
3.2 Review Excel Formulas Created by Cowork
Cowork can create Excel files with working formulas (VLOOKUP, SUMIFS, Pivot Tables), but formulas can be incorrect, such as:
- Formula ranges not covering all data
- Nested IF conditions in the wrong order
- VLOOKUP using Exact Match when it should use Approximate, or vice versa
Pro tip: Ask Cowork to "explain every formula used in this Excel file" and verify they match your requirements.
3.3 Always Review Before Sharing
Documents created by Cowork (Word, PowerPoint, reports) should be reviewed before sending to clients or managers:
- Does the data match the source?
- Are numbers consistent throughout the document?
- Is the language correct? (AI may occasionally make spelling errors)
- Is the tone appropriate for the audience?
Part 4: Speed — Boost Efficiency with Queue, Plugins, and MCP
4.1 Use Queue to Run Multiple Tasks at Once
Instead of issuing one command at a time and waiting, queue multiple tasks consecutively:
- "Organize the Downloads folder by file type"
- "Summarize expenses from all receipts in the Receipts/ folder"
- "Create a PowerPoint summary of this week's work from the Notes/ folder"
Cowork will execute them in order. You don't need to watch the screen — come back and review all results at once.
4.2 Choose the Right Plugin for the Job
Plugins help Cowork perform specialized tasks better. Only enable the ones you need:
| Task Type | Recommended Plugin |
|---|---|
| Data analysis / chart creation | Data Analysis |
| Contract drafting / legal document review | Legal |
| Financial statements / number reports | Finance |
| Content creation / campaigns | Marketing |
| Backlog management / spec writing | Product |
4.3 Connect MCP for Cross-App Workflows
If you need Cowork to work across applications, connect through MCP (Model Context Protocol), such as:
- Google Workspace — Read/create Google Docs, Sheets, Slides
- Slack — Read messages, summarize channels, send messages
- Notion — Read/write Pages, Databases
- Figma — Read designs, export assets
Caution: Only connect the MCP integrations you need, because the more connections you add, the larger the attack surface where AI could access data it shouldn't.
4.4 Set Up Scheduled Tasks for Recurring Work
If you have tasks that repeat daily or weekly, set up a Scheduled Task, such as:
- Every morning: "Summarize important emails that haven't been replied to"
- Every Friday: "Compile action items from this week's meeting notes"
- End of month: "Create an Excel expense summary from Receipts/"
Caution: Scheduled Tasks run even when you're not watching the screen, so define clear boundaries, such as "read only, do not delete files."
5 Things You Should "Never Do" with Cowork
| # | Never Do This | Reason |
|---|---|---|
| 1 | Grant access to your entire Home Directory | AI could read SSH Keys, configs, and browser data |
| 2 | Share files containing passwords/API Keys | Data is sent to the cloud for processing |
| 3 | Trust AI numbers without spot-checking | AI may hallucinate or miscalculate |
| 4 | Let Cowork read files from strangers without inspection | Risk of prompt injection |
| 5 | Set Scheduled Tasks with permission to delete/edit important files | Runs automatically without supervision |
Checklist — Before Every Cowork Session
- Verify that accessible folders contain no sensitive data
- Back up important files before requesting edits
- Write clear instructions specifying What + How + Expected Output
- Use Global Instructions for rules you apply repeatedly
- Check results — spot-check at least 3–5 values
- Review documents before sharing
Cowork is a powerful tool, but "powerful" also means "capable of making big mistakes." Used correctly, it saves enormous amounts of time. Used carelessly, it can create problems harder to fix than doing the work yourself.
— Grand Linux Solution Team
