How to Make Claude Private: Settings and Sharing Boundaries

Make Claude more private by reviewing model improvement, shared chats, saved context and connected applications separately. There is no single switch that makes every conversation invisible to the provider, your organization and outside services. Decide which audience or use you want to limit, then inspect the corresponding control.

When your privacy choices affect shared work, fenbs can keep the team-approved instruction with the project. It does not make Claude private or replace its settings. Save the rule and approved next steps there, rather than copying the private conversation into another service.

Where should you begin the privacy review?

1. Check the signed-in Claude account and selected workspace. Separate personal use from organization work before changing settings.
2. Open Settings and review Privacy. Check the model-improvement control where it applies and inspect the list of shared chats.
3. Review saved context and decide whether the next conversation should use memory or start without it.
4. Review connected applications and remove access you no longer need. Check the source account as well as the connection’s name.
5. Before sharing any answer, inspect its text and intended audience, then choose what authorized work information should be retained.

Choose a clear objective for the review. You might want a conversation absent from normal history, an old shared link disabled or an application disconnected. Write that objective down. “More private” is a direction; “this conversation must not be available through its old public link” is something you can verify.

If you are new to the product, start with our Claude guide. If the workflow takes place in a coding environment, the Claude Code overview helps identify that different interface before you look for the relevant controls. Using another assistant too? Follow the separate ChatGPT privacy setup for that account rather than assuming Claude’s settings carry over.

How do you check the model-improvement preference?

For consumer accounts, Anthropic places the model-improvement control under Settings, then Privacy. Follow its current privacy-setting instructions↗ (opens in a new tab). An organization workspace can have different terms, so ask the administrator when a consumer instruction does not match your account.

Treat the result narrowly: you checked a training preference for a particular account. You have not thereby checked every connected application, chat audience or retention arrangement. Avoid writing “all data collection disabled” in your notes when the action was turning off one form of model improvement.

For work, also check whether the task is permitted at all. A personal preference does not override the client’s conditions or your employer’s handling rule. If the information is not approved for that service, rewrite the task using a harmless example or use the approved environment instead.

What should you do about existing shared chats?

Anthropic’s sharing instructions↗ (opens in a new tab) describe reviewing shared chats in Privacy settings and unsharing a snapshot. Check the relevant conversation and disable sharing when the audience is no longer appropriate. Review the visible conversation, including source details repeated in the answer, before deciding to share again.

Disabling access to a link does not retrieve copies someone already saved. Treat distribution as a separate event: who received the material, where it was forwarded and whether another follow-up is needed. If restricted information reached an unintended audience, use the organization’s incident process rather than relying on the link change alone.

For a future handoff, consider whether a short approved summary will do. You can preserve the decision and next step without distributing the entire prompt history. The goal is to give the recipient enough context to act, not every detail the original conversation happened to contain.

What does incognito change, and what remains?

Anthropic’s incognito guide↗ (opens in a new tab) says these conversations are not saved to ordinary history or memory and are not used for training. They still have retention arrangements, and organization exports can include them. Read those limits before equating the mode with a conversation that nobody else can access.

Choose incognito for a task you are authorized to submit when you do not want a normal ongoing chat record. It is not a workaround for an organization’s restrictions. A sensitive topic may need a different approved service or no AI submission at all, regardless of whether it appears in your sidebar afterward.

Turning memory off is another distinct choice. Anthropic’s memory documentation↗ (opens in a new tab) explains how saved context and individual-chat controls differ. Check both the information already saved and the context you intend to supply next. A fresh conversation can still receive sensitive information through your own prompt.

How should you limit connected applications?

List the services connected to your account and the job each one serves. Open the permissions information before keeping a connection. Anthropic’s connector guide↗ (opens in a new tab) describes how apps provide access under the connected identity. Remove connections that no longer have a useful, approved purpose.

Do the review in both directions. A connected service might return information to the conversation, and an action might send information to that service. Ask what the specific request needs, which account is involved and who can see the destination. A connection you once approved is worth reconsidering when your role changes.

For example, suppose you connected a shared project tool during a short engagement. After the engagement ends, remove unnecessary access instead of leaving the arrangement to memory. For active projects, keep the approved boundary in fenbs Decisions and rules so another assistant does not assume that old access means current permission.

How to do this with fenbs

1. Create a free account and add the project whose privacy boundaries should survive a new chat.
2. Record the person-approved handling rule in Decisions and rules. Include the allowed context, intended services and information that must stay out.
3. Create tasks for remaining checks, such as reviewing a shared link or confirming a work account’s controls. Save the result without private account details.
4. Connect an approved AI app using Integrations. Ask it to read the rules before starting and save only authorized task progress at the end.

fenbs is free for one person, no credit card. It does not monitor Claude settings, import private chats or prevent a person from sharing information elsewhere. If you require technical controls that block a particular transfer, use the organization’s approved system for that requirement.

How can you tell the review is complete?

Compare your results with the original objective. Check the specific switch, link or connection rather than asking Claude whether the account is private. A generated reassurance is not evidence of a setting, agreement or deletion. Record unanswered questions as unfinished, with the person who can resolve them.

Repeat the review when you add a connection, change accounts or introduce a different category of information. Keep the approved rule short enough to follow in an ordinary session. The AI governance hub offers related guides, but the useful outcome here is a concrete boundary you checked and can explain.