Key point. PromptScrubber never calls an AI service. It prepares text locally; any external assistant is a separate decision you make after review.
The concrete problem
Sensitive information in prompts is not limited to names. API keys, access tokens, internal URLs, customer references and confidential project details can all end up in a conversation.
The pressure to get a good answer encourages oversharing. A long paste feels more useful than a short one, even when most of the identifiers are irrelevant.
Placeholders help only if the preparation is deliberate. Reusing the same placeholder for different people, or leaving a token in a quoted line, weakens the result.
Practical steps
- State the question first, then include only the background needed to answer it.
- Remove secrets and internal links manually before any external paste.
- Run the remaining text through local detection and review every highlighted value.
- Use distinct stable placeholders for distinct people or organisations; add missed values yourself.
- Paste the prepared prompt, then restore the answer locally if you need the original references back.
How PromptScrubber helps
PromptScrubber recognises many structured identifiers and lets you mark free-text values manually. It keeps relationships visible without exposing the underlying names, addresses or keys.
Because it runs in the browser with outbound requests blocked, the preparation step itself does not create a new data flow.
Limits and responsibilities
- A model can infer sensitive facts from context even after identifiers are replaced.
- The output is pseudonymized while a mapping exists; it is not a guarantee of irreversible anonymization.
- Do not paste credentials into any external tool. Detection is a safety net, not permission to be careless.
- Scanned PDFs are not processed; DOCX extraction is text-only.
- This guidance is practical, not legal advice or a certification of compliance.
Frequently asked questions
Does the tool check prompts automatically before I paste?
No. You paste text into the local editor, review the result, then copy it. The browser step is deliberate so you remain in control.
What if a secret is not recognised?
Select it in the editor and add it manually. Never rely solely on automatic detection for credentials.
Can I restore the final answer?
Yes, while the tab remains open. Paste the answer into the restore field to reinsert values from the local mapping. Protect any mapping you export.
Do placeholders make a prompt safe enough to share publicly?
Not necessarily. Context can still reveal people or projects. Review the whole text and remove details that are not needed.
Prepare your text
Open the tool: paste, review and anonymize text without uploading it.