Common foundation

Organizing your files for AI

A single root folder, flat work folders, a typology: the condition for AI to find, file and reuse your documents.

CoworkSkillArtifact1 weekend, then 5 min/day

The problem

Your documents are scattered across Downloads, several drives, an institutional Drive and email attachments. The AI can only work on what it sees: a file outside its scope does not exist for it, and an automated task pointing to another drive breaks at the first change of computer.

Step-by-step method

Create a single root folder reserved for the AI (for example IA-NOM). This is the only folder you give Claude access to (Cowork or Claude Code).
Flat work folders, one per major activity: ARTICLES, PRESENTATIONS, SERVICE, ENSEIGNEMENT, BIBLIOGRAPHIE, DONNEES… Avoid deep folder trees.
Three technical folders: _SYSTEME (rules, plans, logs), _A_TRIER (holding area) and 90_ARCHIVE (what is never really deleted).
Write the filing rules in a _SYSTEME/REGLES-CLASSEMENT.md file: what goes where, how to name things, what is sensitive.
Have a typology built: ask Claude to go through a sample and propose file “profiles” (manuscript, conference deck, minutes, activity report…), each with its family, version, sensitivity and target location.
Approve on a dedicated page: Claude displays its proposals in an artifact (Approve / Correct / Discard). Nothing moves before your decision; your corrections become learned rules.
File by copying, not by moving: the original is kept for a few days, then removed after checking. Ambiguous files go into _A_TRIER.
Bring back what is elsewhere: any useful folder located outside the root must be brought back into it.

Known pitfalls

Confidence of a profile: high when the name is recognised and the folder is legitimate, medium when only one of the two is, low when neither is. Only high confidence leads to a direct proposal; the others lead to a question.
A folder that breaks the rules leads to a filing proposal, never to an automatic move.
In a sensitive area, the profile is read from the file name and the folder, without opening the file.
A question that has been settled is never asked again: the answer becomes a learned rule, reread before each run.
A new typology clue (a word in the name, a sender, a format) is tested on the full index before being kept: how many files it would reclassify, and which ones wrongly.
A question of doubt is written in two places: in the run report and in the decision queue, so that it does not get lost.
Before filing an attachment, check that a copy has not already been filed (same name, same size, same date).
File by copying, with a retention period, then a checked removal of the original. A plain move can silently skip a name collision and lose a file.
Archives stay read-only: map and index first, a dated copy to work on, duplicates proposed without moving anything. See Finding anything.

Deliverable

A clean root folder, a rules file, an approved typology and an automatic daily filing of Downloads.

Safeguards

Rule of doubt: Claude asks rather than guesses.
Files containing patient data are flagged as sensitive, their names masked, and they never leave the workstation.
Nothing is ever deleted: it is archived.

Starter prompt

Here is my root folder. Go through it without moving anything. Propose a typology of 10 to 15 file profiles (family, examples, target location, sensitivity) and list the files in doubt. Present everything on a page where I can approve, correct or discard each proposal.

← Back to the triptych