Second Brain Sidekick

Your context, by your side.

How to share context with AI safely

Better context can produce better advice. It does not follow that you should give an assistant your entire vault.

The useful part is usually small: the goal, the constraint, and the decision history that could change the answer. Client names, exact contract values, private messages, access details, and unrelated notes usually add risk without improving the reasoning.

Use four steps before you send context: classify, minimize, redact, and choose the chat mode.

1. Classify the material once

Create three sharing levels so that you do not have to invent a privacy policy in the middle of every conversation.

LevelWhat belongs hereDefault action
ShareableGeneral goals, public facts, working preferencesShare when relevant
Shareable after editingClient situations, financial ranges, internal decisionsSummarize and redact first
Do not sharePasswords, API keys, identity documents, medical records, privileged or contractually restricted materialKeep out of consumer AI chats

The third category is not a challenge to anonymize harder. It is a stop sign. If work is legally or contractually sensitive, follow your employer’s policy and use only approved systems.

2. Minimize before you anonymize

Removing names is not enough if the details are unnecessary. First ask which facts could change the recommendation.

The design studio does not need to paste a client email, its full price list, and last year’s account history. For a new-client pricing question, this may be enough:

Goal: Move toward fewer, higher-value projects this quarter.
Constraint: Keep two long-standing clients on their current rates.
Past decision: A blanket increase cost us a valued client. Do not repeat it.
Question: What is the lowest-risk way to test a new-client rate?

Shorter context is easier to inspect. It also keeps irrelevant history from steering the answer.

3. Redact what remains

Replace details that identify a person or organization while preserving the relationship that matters to the decision.

Before

Acme Dental has paid us $4,850 a month since 2023. Maria said their September budget is frozen. Our contract renews on October 17.

After

One long-standing client is on a mid-four-figure monthly retainer. Its budget is fixed until the next renewal, so its rate cannot change this quarter.

The edited version still contains the decision-relevant facts: the relationship is established, the amount is meaningful, and the timing is fixed. It removes the name, contact, exact amount, and exact date.

Check for identifiers in five places:

  • names of people, clients, products, and locations
  • exact amounts and uncommon dates
  • email addresses, phone numbers, URLs, and account numbers
  • document titles or quoted phrases that can be searched
  • metadata in screenshots, filenames, and copied headers

Redaction lowers exposure; it does not guarantee anonymity. A rare combination of facts can still identify someone. When in doubt, generalize further or do not send it.

4. Choose the chat mode deliberately

Product controls change, so verify the current documentation before relying on a setting.

In ChatGPT, Temporary Chat does not appear in history, does not use or create memories for personalization, and is not used to improve models. OpenAI says a copy may still be kept for up to 30 days for safety, and custom instructions may still apply. See OpenAI’s [Temporary Chat FAQ](https://help.openai.com/en/articles/8914046-temporary-chat-faq) and [Memory FAQ](https://help.openai.com/en/articles/8590148-memory-in-chatgpt-faq).

In Claude, Incognito chats are not saved to chat history or memory and are not used for training. Anthropic says they are retained for a default of 30 days, or according to an organization’s retention policy. See Anthropic’s [Incognito chats guide](https://support.claude.com/en/articles/12260368-use-incognito-chats).

These modes reduce persistence in the product experience. They do not make sensitive information safe to share, override workplace rules, or control what a third-party action or connector does with data.

A 30-second check before Send

Read the context once and ask:

  • Does every line have a plausible chance of changing the answer?
  • Have I removed names, exact figures, contact details, and searchable quotations?
  • Does this contain anything prohibited by a contract, employer policy, or law?
  • Am I using the appropriate history, memory, and training controls?
  • Would I still be comfortable if this text were retained for the provider’s stated period?

If any answer is no or uncertain, stop and revise.

The goal is not to make a perfect anonymous copy of your vault. It is to create the smallest useful brief for one decision.

Use the five-note setup to keep the source material, then share only the slice the current question needs.

Join the Starter Kit waitlist