I trim private data before spending tokens

Start with one task and the smallest useful input

Before sending anything to an AI service, write down the specific help you need. Explaining a database error usually requires the error message and a little surrounding logic, not a complete database export. Reviewing an essay does not require your student number or your classmates’ names.

This first pass protects both privacy and your budget. Smaller requests generally use fewer input tokens, and a focused question makes irrelevant answers less likely. Keep the original material locally and prepare a separate copy for the request. Treat that copy as something you are deliberately sharing outside your device.

Identify everyone who may handle the request

Next, map the route your data takes. A request might pass through an editor extension, an application backend, a relay, and a model provider. Each service may have its own logging, retention, and access practices. An encrypted connection protects data in transit, but it does not mean the services processing the request cannot read it.

If you use MyRelay, include both the relay and the provider serving your chosen model in this check. One key and one budget simplify access; they do not, by themselves, establish a single privacy policy across the whole route. Also check whether your development tools save prompts locally or send diagnostic reports elsewhere.

Check retention and training terms before uploading

Read the current terms for the actual service and account type you plan to use. Consumer chat products and API services can have different rules. Look for whether inputs or outputs may be used for training, how long content is retained, whether people can review it, and what deletion controls exist.

Do not treat an opt-out from training as a promise of zero retention. Security monitoring, backups, and operational logs may have separate rules. If the documentation is unclear, ask support before sending sensitive material. For school or work data, check the institution’s rules too. Access to a document is not automatically permission to upload it to an outside service.

Remove identifiers and secrets from your working copy

Now inspect the material itself. Replace names, email addresses, account numbers, and other identifiers with consistent invented values. Keep relationships intact so the model can still follow the problem. For example, use the same fictional customer name throughout a support conversation rather than changing it in every paragraph.

Remove passwords, API keys, session tokens, and private credentials entirely. Check error messages, screenshots, document comments, and copied headers as well as the main text. A screenshot can reveal a browser tab or account name you did not intend to share. Replacing a name alone is not enough when a distinctive combination of details still identifies someone.

Test the question with synthetic details first

Before using the cleaned copy, try a small request with invented data. For a programming problem, recreate the failure with a few fictional records. For writing feedback, send one non-sensitive paragraph. This tests whether the task really needs more context before you spend money or disclose additional information.

MyRelay lets members see each request’s token usage and cost, which helps make this small-test habit measurable. If you want one place to manage that spending, create a MyRelay account and begin with non-sensitive test requests. A usage view is useful for budgeting, but it is not proof of how request content is stored or deleted.

Review the answer before saving or sharing it

Privacy checks do not end when the response arrives. An answer can repeat personal information from your prompt or infer details you did not explicitly state. Read it before pasting it into a public issue, assignment, shared document, or repository. Remove unnecessary identifiers from the output as carefully as you removed them from the input.

Finally, keep only the prompt history and results you actually need, and use available deletion controls without assuming they erase every retained copy immediately. If you accidentally send a live credential, revoke or rotate it promptly. The dependable habit is simple: share less, verify the route, test cheaply, and review what comes back.