You can check the draft without giving anyone a copy of it.

Three questions decide everything. Does the draft contain personal data, a client's confidential material or an unreleased business fact? If yes to any, the content cannot be uploaded to a service that keeps a copy.

The classification should be explicit and recorded, because the failure mode is not a policy gap. It is a well-meaning person pasting the wrong document into the wrong tab at 11pm.

Once the draft is classified as internal or confidential, the workflow below applies unchanged, which is the point: no special procedure, just a check that does not need the file.

Open the console and paste the text into the content tab, or the whole markdown or HTML file into the document tab. The rules execute in the browser, so nothing is transmitted while the check runs.

You can confirm that rather than trust it: open the browser's network panel, run the check, and watch for requests. Then disconnect networking and run it again. Findings should still appear.

The report tells you what the draft is doing: unsourced statistics, absolute claims, placeholders, structure problems, stock phrasing, hedging density and flat rhythm.

Mechanical findings you fix immediately. An unfinished placeholder, a bare hash link, a heading rank that skips, a statistic with nothing behind it.

Judgement findings need a person and no upload: whether the argument holds, whether a claim will be read as a promise, whether the tone suits the audience.

Do that reading yourself where you have the standing, and note what you changed. The note is what makes the next review cheaper.

For each statistic, write the source beside it in a working copy. Where there is no source, the sentence gets narrowed or cut before anything else happens.

This is the step where confidential drafts most often fail, because the numbers are the part the model invented most confidently and the part nobody can verify without leaving the document.

Keep the source list separate from the draft. When the piece is later shared more widely, the list is what lets somebody else check the claims without the original working copy.

If the draft needs a specialist signature, the review itself implies disclosure, so scope it deliberately: what precisely will be shared, with whom, for how long, and what they are asked to confirm.

Share the findings and the specific sections under review rather than the whole document where that is possible. A reviewer asked to check the eligibility wording does not need the commercial model.

Agree the confidentiality terms before the file moves, not after. A review that starts with an unsigned agreement has already disclosed something.

Keep four lines: what was checked, what was found, what was fixed and who approved. Store it beside the draft rather than in a tool that will outlive the project differently.

That record is what answers the question six months later about whether anyone verified the numbers before publication, and it costs a sentence to write at the time.

It also gives you a comparison for the next draft: the same findings returning means the pipeline has a recurring defect rather than the writer having a bad afternoon.

Sometimes a check genuinely needs a server: transcription, extraction from a scanned PDF, model-based analysis. Those send a copy and create a retention question.

In that case, reduce before you send: extract the pages that matter, redact identifiers, and confirm the provider's retention window in writing. Then record that you did.

The habit worth building is the pause before the upload. Most confidential drafts never need one, and the ones that do can be reduced to something that is safe to send.