Before pasting a customer email into a prompt

A customer email lands in your inbox, and you want help writing a clear reply. Copying the whole thread into an AI prompt feels like the simplest next step. But the thread may contain a phone number, an order reference, a private complaint, and details from earlier messages that the assistant does not need.

A useful starting rule is simple: send the minimum information needed for the task. The goal is not to make every prompt anonymous at any cost. It is to notice which details help the assistant and which details merely expose someone.

§ Keep secrets out entirely

Passwords, API keys, access tokens, private keys, recovery codes, and session cookies do not belong in prompts. These are credentials: someone who obtains them may be able to access an account or system.

If you need help with an authentication error, replace the credential with a label such as REPLACED_API_KEY. Keep the error message and the relevant code structure, but check both for hidden credentials before sending them.

If you accidentally submit a real secret, revoke or rotate it. Editing a local copy or asking the assistant to forget it does not undo the disclosure.

§ Remove personal details that do not change the answer

For a customer reply, the assistant usually needs the issue, the desired outcome, and your tone. It rarely needs the customer’s full name, street address, birth date, payment details, or complete account history.

• Replace names with stable labels such as CUSTOMER_A and SUPPORT_AGENT.
• Remove email addresses, phone numbers, account identifiers, and delivery addresses unless they are necessary and approved for the workflow.
• Summarize medical, financial, or legal circumstances rather than pasting complete records.
• Check quoted replies and signatures, not just the newest message.

Replacing a name alone may not be enough. A job title, small town, exact date, and unusual event can identify a person together. Consider the whole passage, not only obvious identifiers.

§ Use a smaller, safer example

Instead of pasting a complete email thread, try: “Draft a polite reply to a customer whose delivery is three days late. We can offer a replacement or a refund. Do not promise a delivery date.” That gives the assistant enough context without sharing the customer’s identity.

For technical questions, use the smallest code sample or log excerpt that reproduces the problem. Replace real records with invented examples while preserving the structure that matters. If a date or amount affects the answer, change it carefully or explain the relationship in words.

Redaction should preserve the problem, not the private details. Read the revised prompt once before sending it to confirm that it still asks a clear question.

§ Check attachments and automatic context

A prompt is more than the text you type. Screenshots can reveal browser tabs, account names, or private messages. Documents may include comments and tracked changes. Logs can contain authorization headers, database connection strings, and customer requests.

Coding assistants can also send context from files or command output, depending on their settings and permissions. Before using one in a work repository, check which files and tools it can access. Avoid assuming that a file is excluded merely because you did not paste it yourself.

§ Know where the request goes

Vana API provides one endpoint for requests to different model providers. That convenience does not remove the need to review the data path. Check the applicable privacy terms, retention rules, and organizational requirements for both the relay service and the selected provider. Do not assume that every route has identical handling rules.

If you are trying Vana API, create an account and begin with a synthetic example rather than a real customer record. You can test whether a workflow is useful before introducing any approved work data.

§ A short pause before sending

Ask yourself: does this contain a secret, identify a person, or reveal confidential work? Can I remove details without weakening the question? Am I allowed to share what remains with the services involved?

If you cannot answer the last question, stop and check your organization’s policy or ask the responsible person. A smaller prompt is often easier to review, easier to reuse, and sufficient for the task.