AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
Most people start using Claude by opening a chat and typing. That feels efficient, and for the first few days it is. The cost shows up later: chats you assumed were private were eligible for model training, you re-explain your job and your house style in every new conversation, and memory has been quietly filling up with whatever you happened to mention first, including the throwaway test chats.
The rule: before your first real piece of work, set three things in this order: who can learn from your chats, what Claude should always know, and what Claude is allowed to remember. Each one shapes every conversation after it, so doing them later means cleaning up after the fact.
| You need | Notes at time of writing |
|---|---|
| A Claude account | All three settings work on the Free plan, so this costs £0. Pro and Max have them too. |
| Claude on the web or Desktop | Easier than mobile for writing instructions. |
| 20 minutes and a blank document | You will write a short note in Step 2 and keep your own copy. |
If you are on a Team or Enterprise plan, read the guardrails first. Your organisation controls two of these three settings, not you.
Do this first: the switch only governs chats from the moment you change it.
claude.ai/settings/data-privacy-controls.What the switch actually changes, according to Anthropic's privacy centre:
| Setting | What happens to your chats |
|---|---|
| On | New or resumed chats may be kept in de-identified form for up to five years in Anthropic's model training pipelines. |
| Off | New chats are not used for future training. If you delete a conversation, it leaves your history at once and is removed from back-end storage within 30 days. |
| Either | Chats that safety classifiers flag can still be used for trust and safety work. Feedback you send with the thumbs up/down buttons is kept for five years. |
Turning it off does not pull data out of training runs already under way or models already trained, which is why it goes first.
Instructions for Claude apply to every conversation on your account. This is where you stop re-introducing yourself.
If you used Global instructions in Cowork before the new Claude experience reached your account, those instructions have been folded into this box. Open it and check that what is there is still what you want before adding anything.
Here is a prompt to draft the note. Fill in the brackets with your own details.
You are helping me write the account-wide instructions I will paste into Claude's "Instructions for Claude" setting. These apply to every conversation, so they must be short, general and true for all my work.
About me:
- Role and business: [YOUR_ROLE_AND_BUSINESS]
- Who I usually write for: [YOUR_AUDIENCE]
- Style rules I always want: [STYLE_RULES, e.g. British English, prices in £]
- Things Claude must never do: [HARD_LIMITS]
Steps:
1. If any of the inputs above are blank or vague, ask me up to three short questions before drafting. Do not invent details about me.
2. Draft the note under four headings: Who I am, House style, Defaults, Never.
3. Keep it under 250 words. Use plain bullet points.
4. Leave out anything that only applies to one client or one project, and list those items separately underneath as "Better kept in a project".
Before you answer, check that every bullet would hold for any task I give you, that nothing contradicts anything else, and that you have not added preferences I did not give you.
Keep the "Better kept in a project" list for your first Project, which comes after this guide.
On Free, Pro and Max, memory is on by default on the web, Claude Desktop and the mobile app. Claude saves things as individual topics while you chat, such as your role, the people you work with, your working style and ongoing projects. Each Project gets its own separate memory. So the question is not whether to switch memory on, but whether what it holds is accurate.
If you switch memory off, you get two choices. Pause memory keeps what is already saved but stops using it or adding to it. Reset memory deletes everything, including project memories, and cannot be undone.
For a chat you do not want remembered, you have two options. You can click + in the message box and turn Memory off before you send the first message; the chat stays in your history. Or you can open an incognito chat using the ghost icon at the top right of a new chat outside a Project. Incognito chats are not saved to your history and, according to Anthropic, are not used for training even if Step 1 is on.
To audit memory from inside a chat, use this prompt. It needs no inputs, except optionally [ANYTHING_YOU_KNOW_IS_WRONG].
I want to check what you currently remember about me before I rely on it.
1. List each thing you remember about me as a separate line, grouped by topic. Quote what is stored. Do not summarise or improve it.
2. Mark each line as Current, Possibly out of date, or Unsure where it came from.
3. Note anything that contradicts my Instructions for Claude.
4. I already know this is wrong: [ANYTHING_YOU_KNOW_IS_WRONG]. Leave it blank if nothing.
Do not change or delete anything yet. Give me the list first, then ask which items I want changed or removed.
Before you answer, check that you have not added anything that is not actually stored.
Make the corrections in Settings > Memory > Topics, where you can see them take effect.
| Done first | What goes wrong if you skip it |
|---|---|
| Privacy | Every chat until you change it falls under the old setting, and that cannot be undone for runs already in progress. |
| Instructions | Memory starts learning your style from scattered chats rather than from one note you wrote on purpose. |
| Memory review | Wrong facts from early test chats get carried into every conversation from then on. |
Each setting makes the next one work better. Do one and stop, and you get the one-off benefit without the compounding.