AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people open Claude, start typing, and never visit Settings. It feels fine because the chat works, but every default you leave alone is a decision somebody else made for you: what Claude remembers, how it addresses you, where your work lives and what happens to your conversations. The pull of the default is well documented. A widely cited 2003 paper in Science by Johnson and Goldstein compared organ-donation consent across European countries and found far higher consent where donation was the pre-ticked option than where people had to opt in. One secondary summary puts the opt-in rate at roughly five times lower. I could not open the paper itself, so I am not quoting its country-by-country figures.
The rule: for each of five settings, decide on purpose whether the default serves you, and write the decision down. Keeping a default is fine; keeping it by accident is not.
Make a five-row note and fill in one column: "Keep, change, or switch off, and why." It takes about fifteen minutes.
| # | Default | Where to look (at time of writing) | Question to answer |
|---|---|---|---|
| 1 | Memory | Settings > Memory | Do I want Claude building a profile from my chats? |
| 2 | Account-wide instructions | Settings, then the instructions-for-Claude box | What should apply to every conversation? |
| 3 | Loose chats | claude.ai/projects | Which recurring work deserves its own project? |
| 4 | Model-improvement sharing | Settings > Privacy | Am I happy for new chats to be used for training? |
| 5 | Ordinary chat for sensitive topics | Incognito chat | Which topics should never touch memory or history? |
Menu labels move around. Anthropic's help centre has described the privacy toggle as both "Help improve Claude" and "Help Improve our AI models", and the memory article refers to both a new and a legacy layout. If a label differs from the table, look for the closest match and trust the current help page over this one.
Per Anthropic's help centre, memory is on by default for Free, Pro and Max plans. On Team and Enterprise, owners control availability and it is off for members until activated. Claude saves memory as individual topics as you chat. In Settings > Memory you can read, edit or delete each topic, pause memory (keeps what exists, stops new entries) or reset it, which permanently deletes everything and cannot be undone. There are also separate toggles for "Generate memory from chats" and "Search and reference chats".
Decide: keep it on, pause it, or switch it off. If you keep it on, review the topics once. Wrong or stale entries steer answers quietly.
This is the one place to say how you want every answer shaped. Help-centre wording for the field has varied ("What preferences should Claude consider in responses?" in older guidance, "Instructions for Claude" in the current personalisation article), so look in Settings for whichever label your account shows. Keep it to stable things: your role, tone, formatting, what to flag. Anything that changes weekly belongs in a project instead.
Projects are available to everyone; free accounts are capped at five. Create one at claude.ai/projects with "+ New Project", add a name and description, then add instructions and knowledge files. Two details catch people out. Context is not shared between chats inside a project unless it is in the project knowledge, and projects keep their own memory separate from standalone chats. That separation is useful: client work stops leaking into personal chats.
Under Settings > Privacy there is a toggle for using your chats and coding sessions to improve Claude. It applies to the consumer plans (Free, Pro, Max and Claude Code); commercial and API use follows separate terms. Turning it off covers future chats only; data already in a training run that has started stays in it, and conversations flagged by safety classifiers may still be used for safety work. Decide once, knowing the limits.
Incognito chats are available on all plans, are not saved to memory and do not appear in your chat history. They can still be reached by organisation owners through data exports, so on a work account they are not a hiding place.
Priya is a freelance bookkeeper on the Pro plan. Her card reads: (1) Memory: keep on, review topics monthly. (2) Instructions: "UK English, answer first, flag anything that depends on tax year." (3) Projects: one per client, no cross-over. (4) Sharing: off, because client names appear in chats. (5) Incognito: any chat pasting a client's raw figures. Nothing is exotic. It simply stopped being accidental.
Settings > Memory also lets you import from another assistant, and you can tell Claude directly what to remember. Use this prompt in a chat to seed it deliberately.
You are my working assistant. I want to give you durable context so I don't have to repeat myself. Save the following to memory, keeping my wording where possible.
WHO I AM: [ROLE, WHAT I'M BUILDING, WHO IT'S FOR]
HOW I WORK: [E.G. ANSWER FIRST THEN DETAIL, OR CHALLENGE ME WHEN I'M VAGUE]
STANDING PREFERENCES: [TONE, FORMATTING, THINGS TO ALWAYS FLAG, THINGS NEVER TO ASSUME]
CURRENT PROJECTS: [TWO OR THREE NAMES]
OFF LIMITS: [TOPICS OR DETAILS YOU MUST NOT STORE, E.G. CLIENT NAMES, HEALTH, FINANCES]
Rules:
- If anything I ask later conflicts with this, tell me about the conflict and ask which wins. Do not choose silently.
- If any field above is blank or still in square brackets, ask me for it. Do not invent it.
- Do not store anything from the OFF LIMITS line.
Before replying, check that you have not added details I did not give you. Then list exactly what you saved, one line each.
Fill in every square-bracket field; delete OFF LIMITS if you have nothing to exclude. Afterwards, open Settings > Memory and confirm the saved topics match the list.