How to Redact a Document Before an AI Rewrite
Sometimes you want an AI tool to improve a document that contains things that should not leave your control, and the tool you have runs in the cloud. The safe move, if you cannot keep the whole thing local, is to redact the sensitive parts before the text is sent, rewrite the safe version, and put the sensitive details back afterward. Done carelessly this is worse than useless, because a half-redaction leaks the very thing you were protecting. Done properly it is a reasonable fallback.
Decide What Actually Needs To Go
Redaction only works if you are clear on what is sensitive. The usual list: names of people and clients, specific identifiers, financial figures tied to a party, unreleased plans, anything personal, and anything covered by an agreement. Personal data in particular carries legal weight under regimes like the GDPR, which is a reason to keep it off a third-party service. The test is whether this detail, seen by someone outside, would cause a problem. If yes, it goes before the text is sent.
What often does not need redacting is the structure and the generic prose. You can usually rewrite the shape of a document, the phrasing and tone, without the engine ever seeing the specific figures. The reasoning overlaps with what to keep local and what to send to the cloud.
Replace, Do Not Just Delete
The mistake that breaks a redaction-then-rewrite workflow is deleting the sensitive parts, because the rewrite then produces text with holes, and stitching it back together reintroduces errors. The better method is to replace each sensitive item with a clear, consistent placeholder the model can preserve, then swap the real values back after.
Before:
Our contract with Acme Corp for 240,000 euros expires on 3 March, and we plan to propose a 15 percent increase before the board meeting.
After:
Our contract with [CLIENT] for [AMOUNT] expires on [DATE], and we plan to propose a [CHANGE] before the board meeting.
The model rewrites the structure and tone around stable placeholders, and you restore the real values locally afterward. Keeping placeholders intact is exactly the fact-preservation discipline a good rewrite instruction enforces.
Tell The Tool Not To Touch The Placeholders
A rewrite can helpfully “fix” a placeholder into something that looks nicer and breaks the swap-back. So the instruction has to pin them. A Wrivio Context for a redacted rewrite could say:
Rewrite this to be clear and professional. Treat anything in square brackets like [CLIENT] or [AMOUNT] as fixed tokens: keep them exactly as written, do not rephrase, expand, or remove them. Keep every date and figure exactly as written, and do not add anything I did not include.
Press Ctrl+Shift+Space, paste the redacted version, and check the diff: every placeholder should survive untouched. If one changed, fix it before restoring the real values.
The Fallback Is Worse Than Just Staying Local
Be honest about what redaction is: a workaround for when you are forced onto a cloud tool with sensitive material. It is error-prone, because one missed name defeats the whole exercise, and it adds manual steps on both ends. If you have the option, running the rewrite on a local model avoids the entire problem, because the sensitive text never leaves the machine and there is nothing to redact. The guidance for sensitive text is simply to keep it local, as in is it safe to paste work emails into ChatGPT and the risks of pasting company NDA material into ChatGPT. For handing work to automated systems, see redaction before you hand work to an agent.
Common Questions
What should I redact before an AI rewrite?
Strip anything that would cause a problem if seen outside: names of people and clients, specific identifiers, financial figures tied to a party, unreleased plans, personal data, and anything under an agreement, while generic prose and structure can usually stay.
Should I delete or replace sensitive parts?
Replace them with clear, consistent placeholders like [CLIENT] or [AMOUNT] rather than deleting them, because deletion leaves holes that are error-prone to stitch back, while placeholders let the model rewrite around them so you can restore the real values after.
How do I stop the AI from changing my placeholders?
Instruct it to treat bracketed tokens as fixed and keep them exactly as written, then check the diff after the rewrite so every placeholder survived untouched before you swap the real values back in.
Is redaction better than using local AI?
No, redaction is a fallback for when you are forced onto a cloud tool; it is error-prone because one missed detail defeats it, so running the rewrite on a local model is safer since the sensitive text never leaves your machine.
Download Wrivio for Windows to rewrite sensitive text on a local model and skip the redaction dance entirely.
Read Next
Indirect Prompt Injection, Explained for People Who Just Write Emails
Security researchers reported prompt injection moving from theory to operational attacks in 2026. What it is, why it is unsolved, and when it touches ordinary writing work.
Air-Gapped AI Writing: What It Really Means
Air-gapped sounds absolute, but most tools that claim it are not. What air-gapped actually requires, how local AI gets close, and how to verify the claim.
What to Keep Local and What to Send to the Cloud
You do not have to pick one engine forever. A simple rule for which drafts stay on your machine and which can go to the cloud, with a worked line.
How to Structure a How-To for AI Answers
AI engines love step-by-step content, but only if the steps parse cleanly. How to structure a how-to so an engine can lift the steps in order without garbling them.
This article is filed underPrivacy & Compliance, which has 114 articles.