How to Write Claude Cowork Global Instructions That Actually Hold (No Template Needed)

By Branislav Koist · Updated

Good Claude Cowork global instructions are short and about you: who you are, how you want answers, and the few rules that must hold in every conversation. Project details belong elsewhere. Each rule should also name something you can check, because Anthropic describes these instructions as guidance Claude should consider, not as a lock.

Below: why there's no template here, where the setting lives today, what to put in and leave out, and how to make a rule checkable.

Why a template won't write them for you

Many people search for a ready-made template. But look at what Anthropic's Claude Academy suggests putting into global instructions: who you are and what you do, the shorthand you use, and how you like output delivered. All three are about you. A template written for someone else describes someone else.

A borrowed list also tends to be long, and long lists lose rules and breed contradictions, as you'll see below.

The Academy gives better advice than any template: your global instructions don't need to be complete on day one, and the corrections you keep giving Claude are the best candidates. They fit your work as no stranger's list can.

Where Claude Cowork global instructions live today

In what Anthropic calls the new Claude experience, chat and Cowork are one. There, the Cowork Global instructions setting is part of Instructions for Claude in Settings > General and applies to every conversation, not only to Cowork. The change is rolling out in stages: if your message box still shows separate "Chat" and "Cowork" options, you don't have it yet. Until then, Anthropic's Cowork help article describes the older route: open Settings > Cowork and click "Edit" next to Global instructions.

Menu and setting names were checked on 12 October 2026. The Claude desktop app changes them from time to time, so if you can't find a setting, search for it by name.

Anthropic calls this an account-wide setting, so whatever you write there follows you into every conversation, including quick questions that have nothing to do with your projects. After you save, close the setting and open it again: Cowork's changelog from 15 September 2026 lists a fix for Instructions for Claude appearing blank right after saving.

Global instructions are only one layer. The Academy sums it up: "Global instructions cover you. Projects cover what you're working on." For a full map of which layer reaches what, including CLAUDE.md files, read Cowork global instructions vs. CLAUDE.md: what goes where.

What belongs in global instructions

One question decides it: would this line be true in every conversation you will ever have with Claude? If yes, it may belong. If it's true only for one client, one project or one week, it doesn't.

Lines that usually pass that test:

  • Who you are, in a sentence or two. Your role and the kind of work you do.
  • Your shorthand. The abbreviations and names you use, so Claude doesn't have to ask.
  • How you want answers. Language, length, format and tone.
  • A handful of rules where a mistake is expensive. Think about which mistakes have cost you real time or money: a step that couldn't be undone, a number nobody checked, a "done" that wasn't done. Those earn a place everywhere; most other rules don't.

What to leave out

  • Project details. Client names, folder layouts and deadlines belong in project instructions or in a CLAUDE.md file in the project folder. In the global layer, they leak into unrelated conversations.
  • Anything that changes often, and long explanations. Anthropic's best-practices page for Claude Code lists both among the things to leave out of CLAUDE.md, and the same logic fits here. A rule needs one sentence, not its history.
  • Wishes Claude already shares. "Be accurate" or "do good work" adds length without changing anything. Anthropic's documentation recommends instructions that are "concrete enough to verify" instead.
  • Passwords, card numbers and API keys. Never put a secret into instructions. You sign in to services yourself.
  • Emphasis on every line. Anthropic's advice is to add emphasis such as "IMPORTANT" only to the one line Claude keeps skipping. If many lines are marked as important, none of them stands out.

Why shorter holds better

On the pages we checked, Anthropic gives no length limit for Instructions for Claude. It does give one for CLAUDE.md files in Claude Code, and the reasons carry over, because the global layer loads into every conversation too. The documentation says that the more specific and concise your instructions are, the more consistently Claude follows them, and it sets a target of under 200 lines per file, because longer files reduce adherence. The best-practices page adds: if Claude keeps doing something despite a rule against it, the file is probably too long and the rule is getting lost.

Length also breeds contradictions. If two instructions contradict each other, the documentation warns, Claude may pick one arbitrarily.

The best-practices page offers a test for every line: "Would removing this cause Claude to make mistakes?" If not, cut it. In our own setup, the personal file that loads in every Claude Code session holds six short points. The detailed rules live in a separate file on disk, and the first of those six points tells Claude to read it.

Why a rule without a receipt doesn't hold

Written instructions guide Claude; they don't force it. Claude Code's documentation says Claude treats CLAUDE.md files as context, not enforced configuration, and points to hooks and settings for anything that must happen regardless. The help pages about Instructions for Claude that we read describe no check either.

So a rule like "always give the correct time" is a wish. It depends on Claude getting it right, and when it doesn't, the wrong answer reads just as confidently as a right one. Anthropic's best-practices page names the gap: without a check it can run, "looks done" is the only signal available.

Here is a real case from our setup, checked while writing this guide. On a Windows PC, one standard way of asking a command window for the time in a named time zone simply ignored that time zone. It printed 09:32 when the local time was 11:32. Any sentence built on that reading would have been two hours wrong and would have looked perfectly normal.

Our rule about times doesn't say "be careful with times". It says where a time must come from: one measuring tool, never memory. That tool knows about this problem, says so in its output and reads the time in a way that handles the time zone. It also writes every measurement into a log, so a guard script can later check that a measurement existed before a time was written into the change log.

That is the pattern behind a rule that holds:

  1. Name the proof. Not "check your work", but what the proof looks like: a file name, an output, a link, the source of a number.
  2. Make the proof land somewhere. A reply in chat scrolls away. A line in a file stays where you, or a script, can look.
  3. Let something other than the AI look. Your own eyes work. A script works without getting tired.

Global instructions can carry the first point on their own. The second and third need files, and Cowork reaches files only in folders you've connected.

How to write and test them, step by step

  1. Start with a few sentences. Who you are, what you do, how you want answers. The Academy's advice is not to overthink the first version.
  2. Collect corrections for a week. Each time you correct Claude the same way twice, write it down. That's a candidate line.
  3. Run the removal test on every line. Then rewrite each remaining rule so it names its proof. If you can't say what the proof would look like, the line is still a wish.
  4. Test in a fresh conversation. Anthropic's advice for CLAUDE.md fits here too: watch whether what Claude does actually changes.
  5. Prune when something goes wrong. A rule that keeps slipping is a reason to shorten the rest, not to add a louder copy.

If rules still slip, read why Claude ignores your CLAUDE.md and how to check when your AI says "done".

Claims Need Receipts adds the part that global instructions can't carry, in Claude Cowork and Claude Code, for people who aren't programmers: measurements and decisions leave a trace in plain files on your computer, and guard scripts print a warning where a receipt is missing. To be fair, the scripts check 92 of its 216 rules; the rest are about how the AI acts in conversation, and the AI holds those on its own.

FAQ

Is there a Claude Cowork global instructions template?

Anthropic's Claude Academy offers a short starter: a few sentences about your role, your main work, your shorthand and how you like updates delivered. Beyond that, a template written for someone else describes someone else. Start with a few sentences about yourself, then add the corrections you find yourself giving again and again.

Where do I find Claude Cowork global instructions?

In the new Claude experience, they are part of Instructions for Claude in Settings > General, and they apply to every conversation. Without the new experience, Anthropic's help article says to open Settings > Cowork and click "Edit" next to Global instructions. After saving, open the setting again to confirm that your text is there.

How long should Claude Cowork global instructions be?

The help pages we checked give no limit for this setting. For CLAUDE.md files, Anthropic sets a target of under 200 lines and says longer files reduce adherence. Global instructions reach every conversation, so keep them much shorter: who you are, how you want answers, and a few rules where mistakes are expensive.

What is a guard script, and do I need to write one?

No. A guard script is a small program that checks files on your disk (a date, a number, a log entry) and prints a warning when a receipt is missing. In Claims Need Receipts you don't write or edit any of them: your AI reads the guide, asks you seven questions, shows what it will add and installs everything only after you say yes. The one requirement is Python 3.10 or newer on your computer; if it's missing, your AI walks you through installing it.

Sources

All accessed 12 October 2026.

Last updated: 12 October 2026

Claude is a trademark of Anthropic. Claims Need Receipts is an independent product and is not affiliated with, endorsed or sponsored by Anthropic.