Give AI a folder

If you are the kind of person who saves things on your desktop, I have good news: you may be better prepared for the next phase of AI than you think.

Most of us learned to use AI through a blank chat box. Open a conversation. Paste in a document. Explain the project again. Ask for an output. Copy the answer back into Word, Excel, or wherever the actual work lives.

That is fine for a quick question. It gets old fast when the job includes six PDFs, a spreadsheet, last month's proposal, and a folder of photos named some variation of final-final-2.

The more interesting shift is happening on the desktop. Tools like ChatGPT Work—and Codex for more technical project work—can operate inside a local folder you choose. OpenAI's current guidance says you can limit access to the files a task needs, and that local chats, files, and outputs remain on the computer unless you move or share them.

That sounds like a product feature. I think it is actually a workflow change.

The folder is the context

A folder already contains a rough model of the work. The source files are together. Previous versions are nearby. The output has somewhere to go. You do not have to rebuild the entire backstory every time you start a chat.

This matters because many practical office tasks are not really questions. They are small production jobs:

• Compare this month's estimate with the last approved version.
• Turn these site notes and photos into a draft closeout packet.
• Review a new matter folder and list the documents still missing.
• Clean this spreadsheet and create a client-ready summary.

The useful shift is from asking AI a question to handing it a bounded piece of work.

Bounded is important. “Look through my computer and find something useful” is a terrible starting brief. One folder gives the task a visible edge. It makes the sources easier to inspect, the output easier to find, and a mistake easier to contain.

Start with one boring deliverable

I would not begin with your entire client archive, accounting folder, or shared drive. Start with copies of low-risk files and a deliverable you already know how to judge.

A useful first experiment has four parts:

1. One working folder. Include only the source files needed for the task.
2. One defined output. Ask for a revised spreadsheet, comparison memo, checklist, or draft packet—not “insights.”
3. One output location. Keep generated work separate from the source files so nobody confuses a draft with an approved record.
4. One review pass. Check facts, calculations, omissions, formatting, and anything that could affect a customer or a decision.

Imagine a contractor preparing a weekly job update. The folder might contain the estimate, change notes, site photos, and a simple status template. The assignment is not “manage the project.” It is “prepare a draft update using this template, flag contradictions, and save it in the review folder.”

Or take an architecture studio assembling a proposal. The AI can compare the request, prior proposal language, team bios, and project list, then prepare a first draft. A person still decides what is true, what is relevant, and what the firm is willing to promise.

This is also where the boring safeguards matter. Local does not automatically mean safe. Product settings, network access, workplace rules, backups, and the sensitivity of the files still count. The folder is a useful boundary, not a privacy force field.

The useful version is smaller

The goal is not to give AI your whole digital life. It is to remove the repeated handling around one familiar job.

Pick a task you do often. Give it a clean boundary. Define what finished looks like. Then measure whether the result saved time after review—not before it.

If it works, you have something more valuable than a good chat. You have the beginning of a repeatable system.