A useful prompt starts with a data boundary

A prompt often brings together more information than the immediate task needs. A request to summarize a customer conversation may include names, contact details, internal account identifiers, and unrelated history. A useful starting point is to separate the task from the identity of the people involved.

List the facts the model needs to perform the task. Then examine the remaining fields. Some can be removed. Others can be generalized, such as replacing a precise date with a month when precision is unnecessary. Fields that must remain distinguishable can be represented by consistent tokens.

A data boundary is a clear statement about where information may travel. It should cover the application, model provider, storage, logs, support tooling, and any other service receiving the request. The boundary also needs to account for what comes back in the response.

Tokenization can preserve relationships in a prompt. Replacing the same customer with the same token lets the model follow the conversation without receiving the original name. If reconstruction is required, the mapping needs its own access rules and expiry. Sending the mapping alongside the protected prompt would defeat that separation.

Context matters. An organization name, an unusual job title, and a precise event date may identify someone together even when an email address has been removed. Testing should include these combinations and the document formats the organization actually uses. A transformation should be reviewed for both residual exposure and loss of usefulness.

Make the decision visible to the user when uncertainty matters. A review step can show which categories were changed and give an authorized person a way to resolve an ambiguous field. A practical privacy workflow also defines when a request must be blocked.

The Shaige website demonstration illustrates this flow with predefined synthetic text. It does not measure detection accuracy, accept confidential input, or call an external model. A production implementation needs evaluation against the intended use case and a review of its operational controls.