Every time you send a prompt in Cowork, Claude stitches together three things: your prompt, a summary pulled from memory, and your global instructions. That last piece is the one you control completely. It is a short text block that Claude reads before every request, and it is the cleanest way to teach Claude who you are without retyping it in every conversation.
This module covers what to put in global instructions, what to leave out, and how to avoid the most common mistake — turning them into a dumping ground that quietly makes every response worse.
Who this is for
You’ve done setup, turned on memory, and run a couple of sessions. You’ve noticed that you keep typing the same context — “I’m an investment analyst, save spreadsheets as Excel, write in markdown” — at the top of every prompt. Global instructions are the fix.
What actually belongs in global instructions
Four categories, roughly in order of usefulness:
- Your role and industry. One sentence. “I’m an AI consultant for investment management firms.” This reframes ambiguous requests the way a good colleague would.
- Preferred output formats. “If you’re creating a text file, use markdown. For spreadsheets, use Excel. For slides, use PowerPoint.” Without this, Cowork guesses — and usually guesses Google Docs.
- Hard rules. “Never delete a file without confirming with me first.” These matter because they are firm. Memory is not reliable enough to enforce a rule; global instructions are.
- Firm-specific vocabulary. Acronyms, internal product names, the way your shop uses jargon. Saves Claude from asking or guessing.
That’s it. Resist the urge to add more.
What does not belong there
- Project-specific context. “We are writing the Q2 letter” is a project instruction, not a global one. It will pollute every unrelated chat you start tomorrow.
- Long style guides. If your firm has a real style guide, it belongs in a skill or a project file. Global instructions are too small a surface for it.
- Reference data. Price lists, client rosters, org charts. These are files. Upload them to a project.
- Anything you only want sometimes. If a rule only applies to a particular kind of work, it is not global.
A good test: would you want this rule applied to a chat where you’re asking Claude to help plan a kid’s birthday party? If not, it is not global.
Where to find it
In Cowork:
- Click the gear icon or open Settings.
- Go to the Cowork tab.
- Click Global instructions. You’ll see the note: Instructions here apply to all Cowork sessions. Use this for preferences, conventions, or context that Claude should always know.
- Hit Edit, write your block, and Save.
You can come back and change it at any time. Edits apply to the next prompt you send.
A worked example
Here’s a reasonable first draft for an investment analyst:
I'm an AI consultant and trainer focused on the investment
management industry.
Output formats:
- Text files: markdown
- Spreadsheets: Excel
- Slides: PowerPoint
- Occasionally I'll ask you to post to Notion
Rules:
- Never delete a file without confirming with me first.
- Dollar figures in thousands, no decimals, unless I specify.
Context:
- My newsletter is "Future-Proof Your Career with AI" (substack).
- My website is khehy.com; I reference it often.
Three things to notice. First, it is short. Second, every rule is general enough that it applies to any Cowork session. Third, the “context” lines give Claude information it would otherwise have to ask about every time.
Verifying it actually works
Open a new Cowork chat and ask something that should be shaped by your instructions:
What are my preferences for text files?
If your instructions are being read, Claude should answer specifically — “markdown” in the example above — without you having to explain. If it asks or guesses wrong, check that you saved the instructions and that you’re in Cowork (not Chat).
Common mistakes
- Writing a novel. Long instructions crowd out the actual prompt and can make responses worse, not better. Keep it tight.
- Putting project-specific context there. It leaks into every chat. Use a project for anything scoped to a body of work.
- Assuming it applies to Chat. It does not. Chat has separate personalization.
- Trusting it for precise quoting or safety rules. Global instructions are strong but not bulletproof. For anything critical — compliance language, legal disclaimers — verify every output anyway.
- Leaving stale rules. If your role or workflow changes, update it. Old instructions silently distort everything.
Global instructions vs memory vs projects
A quick mental map:
- Global instructions: authored by you, injected verbatim, every Cowork session. Firm and durable.
- Memory: inferred from your chats by Claude. Soft and leaky. Good for background, bad for hard rules.
- Projects: scoped to a body of work. Files plus project-specific instructions plus a rolling memory. Use these for anything that doesn’t belong in every chat.
If a rule should apply everywhere, global. If it should apply to a specific body of work, project. If it should be inferred over time, memory.
Next up
Global instructions handle the “who you are” problem. Connectors handle the “what you can reach” problem — hooking Cowork up to your email, calendar, Notion, Google Drive, and the other tools where your actual work lives.