Prompting without leaking

Most tasks that tempt you to paste sensitive data do not actually need it. The model needs the structure of your problem, rarely the identifying details. Redacting is a skill, and it takes seconds once practiced.

The three techniques

1. Placeholders

Replace identifying details with neutral tokens before pasting, and restore them afterwards. "Draft a payment reminder to [CLIENT], invoice [NUMBER], amount [AMOUNT], 30 days overdue." The model writes exactly as good a letter, and the identifying details stayed home. One caution: redaction reduces risk, it does not eliminate it. Reread what remains for indirect identifiers (a job title plus a city can identify a person) and for confidential detail that survived the redaction.

2. Abstraction

Describe the situation one level up. Instead of pasting the angry email from a named customer, write: "A long-standing customer is angry about a delayed delivery and threatens to leave. Draft a de-escalating reply offering a gesture." You typed five seconds longer and shared almost nothing.

3. Synthetic examples

When you need help with a format or a formula, recreate the problem with fake data: a spreadsheet issue with invented numbers, a code bug reproduced in a toy snippet. The answer transfers; the real data stays home.

Two smaller habits complete the picture. Check your AI tools' settings once: many offer a toggle to exclude your conversations from training, and turning history off where offered reduces what is retained. And treat chat history as a data store: a year of pasted work fragments is itself a leak waiting for an account compromise.