How AnonymFiles works, step by step

The journey

Seven steps, from your folder to the AI's answer.

1. 1Drop in the folder

   Drag a whole folder into the window, subfolders included, or choose it. To try without a real document, the app offers a built-in sample folder.

   • Read: PDFs, text or scanned (local OCR), Word .docx documents (and .doc on Mac, converted locally), .eml and .msg e-mails, .xlsx, .xls and .csv spreadsheets, .pptx and .ppt presentations, .jpg, .png and .tif images, and .txt and .md files, up to 50 MB each (10 MB for text and spreadsheets).
   • Flagged and excluded, never copied as is: encrypted or damaged PDFs, .doc files on Windows, archives and e-mail attachments.
   • Your originals are never modified. The app works on copies, in a private workspace on your computer.

   

2. 2Detection finds what identifies people

   Detection starts as soon as you drop the folder, and you follow its progress document by document. Two models installed on the computer find the names of people, organizations, addresses, private dates and identification numbers; rules find e-mails, French-format phone numbers and IBANs from nine European countries.

   • Each item becomes a stable label: [PERSONNE_1], [ORGANISATION_1], [ADRESSE_1]… The same person keeps the same label in every document of the folder, and several variants of a name can be linked to the same label.
   • On a Mac, the models are included in the app: no connection. On Windows, they are downloaded once, on the first processing run, without any document text.
   • When in doubt, nothing leaves: if the model fails to load or fails, export stays blocked, with no silent fallback to partial detection.

   

3. 3Compare the original and the copy

   Each document is shown side by side: the original on the left, the copy on the right, exactly as the AI will receive it. A “copy only” mode lets you read it in full.

   • You see precisely what was replaced, and by which label.
   • A copy can still give someone away through an indirect detail: a file number, a place, a very precise date. This is the step where you spot it; the profession section further down lists the typical cases.

   

4. 4Mask what's missing, in one click

   Select a missed passage in the original: an amount, a file number, a reference, a technical passage. It becomes a label in the copy, and the rule applies to the whole folder, not just this document.

   • Amounts are not detected automatically: this is where you mask them, if the AI doesn't need to see them.
   • Meaningful labels: click a label to rename it, for instance [PERSONNE_CLIENTE]. The name stays in square brackets, so that putting names back works.

   

5. 5Approve each document, then export

   Nothing is exported without your approval, document by document. If the original, the rules or the copy change afterwards, the document must be approved again. Export then creates a new folder, wherever you choose; an existing export is never overwritten.

   • Copies for the AI, with their labels. A text PDF stays a PDF, with its original layout: each masked passage becomes a black box carrying its label, and the text underneath is actually deleted (an image that overlaps a masked passage is removed). A Word file gives a .docx redacted in place, layout preserved, plus a text copy (Markdown) for the AI: body, headers, footers and notes are processed; comments, tracked changes, images and properties are removed. .txt and .md files stay text.
   • Optionally, a read-only PDF copy of each Word or text file: the black redaction of the preview, with readable labels. It is generated from the already processed text: replaced values are not hidden under the boxes, they are simply no longer there. The Word layout is not reproduced.
   • Neutral file names (piece-0001…), the subfolder hierarchy preserved, and an index of exported and excluded documents, without the original paths.
   • Hidden layers removed: metadata, annotations, forms and links.
   • No mapping table in the export. What lies behind each label stays in the private workspace, on your computer.

   

6. 6Work with the AI of your choice

   After export, the “Copier la consigne IA” button copies the instructions to paste at the start of the conversation: keep the labels as they are, and repeat the folder reference, of the form [DOSSIER_…], in the answer. The exported folder also contains these instructions, in an AGENTS.md file.

   • You give the copies to the AI of your choice: ChatGPT, Claude, Le Chat, Copilot… The app itself talks to no AI: you choose what to send, and where.
   • Whether the instruction is followed depends on the tool: always reread the answer.
   • Prefer a business plan that does not train its models on your data.

   

7. 7Put the real names back

   In the app, « Réinjecter un résultat » (put a result back) accepts the AI's answer pasted as is, or a .txt or .md file of up to 10 MB, or a .docx or .pdf file of up to 50 MB. The [DOSSIER_…] reference finds the right mapping, even after a restart, without reopening the original folder.

   • A preview shows the text with the real names; you save it to a new text, Word or PDF file, on your computer. This replacement uses no model and no connection.
   • The limits, frankly: in a restored PDF, the value is rewritten in Helvetica, not in the original font. A label comes back in a representative form of the name, and grammar is not adjusted. A missing or unknown reference, or one mixing several exports, blocks processing; unknown or modified labels are flagged, without guessing their content. Reread the preview before saving.