Set it up properly

1. Draft your Instructions for Claude

Instructions for Claude apply to every conversation, so you stop repeating yourself. Let Claude draft them for you.

You are helping me write the account-wide "Instructions for Claude" that will apply to every conversation I have with you.

About me:
- Name and role: {{name_and_role}}
- Organisation: {{organisation}}
- Who I usually write for: {{audience}}
- Things I dislike in AI answers: {{dislikes}}

Write the instructions in the first person, ready to paste. Cover: who I am; UK English spelling; UK date format (8 November 2026); £ for money; lead with the answer, then the detail; plain language with no hype or filler; say when you are unsure and how I can check; ask one clarifying question when a request is ambiguous; and my dislikes.

Constraints: under 150 words, short lines, no headings. Output only the instructions text.

• Your name and role, for example: Priya Shah, HR manager
• Your organisation, for example: a 120-person housing association in Leeds
• Who you usually write for, for example: line managers and tenants
• Things you dislike in AI answers, for example: US spellings, emoji, "I hope this helps", long introductions

Where to paste: A new chat

What you should get: A block of under 150 words in the first person, covering UK spelling, dates, £, tone and your dislikes. Check it sounds like you and contains nothing you wouldn't want applied to every chat.

Follow-ups

• Make it shorter: under 100 words, keeping only the rules that change the output most.
• Add a line telling Claude to format any email draft with a subject line first.

2. Save your instructions

Paste what Claude drafted into your account settings. If you used Global instructions in Cowork before, they now live here.

1. Click your initials in the lower left, then Settings
2. Open Your Claude In the older layout this is Account.
3. Paste your text under Instructions and save it

3. Check your memory

team, enterprise: Your owner controls memory. In organisations created on or after 31 October 2026 it starts on for each member; in older ones it starts off.

Claude saves memories as topics you can read and edit. One wrong "fact" about you quietly skews every answer.

1. In Settings > Your Claude, select Manage next to Your memory Older layout: Settings > Memory.
2. Open each topic and edit or delete anything wrong
3. Leave "Include sensitive topics in memory" off unless you need it
4. Check the "Search and reference chats" toggle is on Paid plans only. It lets Claude find earlier chats when you ask.

4. Switch off training for client work

team, enterprise: This toggle is for Free, Pro, Max and Ultra. Team and Enterprise are commercial plans covered by separate terms: check with your admin.

If you're opted in, de-identified chats can be kept for up to 5 years for training. Turn it off before you paste in client, customer or staff information.

1. Open Settings > Privacy
2. Turn off "Help Improve our AI models"
3. Don't thumbs-rate chats with sensitive content Feedback is kept for five years.

5. Turn on file creation and choose permissions

enterprise: On Enterprise, owners manage capabilities in Organization settings > Capabilities, and the permission setting defaults to Manual.

File creation powers Word, Excel, PowerPoint and PDF output, and skills. The permission setting decides whether Claude asks before it acts.

1. Open Settings > Capabilities
2. Check "Code execution and file creation" is on Called "Create and edit files" in the updated layout.
3. In the message box, set the permission setting to Manual if you want Claude to ask before acting Auto is the default on Pro, Max, Ultra and Team.

6. Prove it worked

A quick test that shows whether your instructions and memory are being used.

Quick test of my settings. Without asking me for more detail, write a 120-word email to {{client_name}} confirming that our meeting has moved to {{new_date}} and that the agenda is unchanged.

Then, under the heading "Settings check", list in 3 bullets which of my instructions you followed (spelling, date format, tone) and anything you know about me from memory that you used.

• A client or colleague, for example: Tom Reid at Fernwood Builders
• The new date and time, for example: Thursday 13 November at 10:30

Where to paste: A new chat (not Incognito)

What you should get: A short email in UK English with the date written the UK way, no hype or filler, followed by a settings check that names your rules. If the spelling or tone is off, go back to Settings > Your Claude and tighten your instructions.

Follow-ups

• Now write the same message as a two-line text instead of an email.
• What else would be useful to add to my instructions, based on this result?

Quick check

1. Where do you switch off model training on a Pro plan?

   • Settings > Usage
   • Settings > Privacy > Help Improve our AI models (correct)
   • Customize > Skills
   • Settings > Capabilities

   The training toggle is in Settings > Privacy.

2. If you're opted in to training, how long can de-identified chats be kept?

   • 30 days
   • 1 year
   • Up to 5 years (correct)
   • Forever

   Anthropic's retention article says up to 5 years for opted-in chats.

3. What does an Incognito chat do?

   • Deletes all your memories
   • Skips memory and chat history for that conversation (correct)
   • Turns off web search
   • Shares the chat with your admin

   Incognito chats aren't saved to memory or history, and Claude won't search them later.