Your Files Are the Agent

The session I get asked for most this year has a strange shape. Nobody asks for it by name. They describe a feeling.

"We have everything. Ten years of proposals, every contract, the pricing grid, the client emails. And every time someone needs to write a new proposal, they start from a blank page and a vague memory of the last one."

That's the feeling. The training that fixes it is built around one tool and one idea. The tool is Claude Cowork. The idea is that your files are the agent.

What Cowork actually is

Claude Cowork started in January as a research preview on the Mac, for Max subscribers only. In April it went generally available on every paid plan, on Mac and Windows, and it has been spreading to the web and mobile since. If you use Claude for chat and you haven't opened Cowork, you've been using half the product.

The difference is simple to state and hard to appreciate until you see it. Chat advises. Cowork does. You grant it access to a folder on your machine, and it reads, writes, edits and creates files in that folder. It builds an Excel sheet with working formulas, not a description of one. It edits a draft in place. When the work is large, it splits it into pieces and runs sub-agents in parallel, so twenty documents don't get read one at a time.

That's the part people expect. The part that turns it from a helper into an agent is what sits on top: Projects, Skills, Plugins, and scheduled tasks.

Start with the folder, not the prompt

Every training I've done with Cowork begins the same way, and it's the part participants resist.

We don't write a prompt. We build a folder.

One folder, on the participant's own machine, with the material that actually defines how their team works. Three good proposals, not thirty. The pricing grid. The brand guide, if it exists, and the two emails that explain the unwritten rules if it doesn't. The template they wish everyone used. The examples of what went wrong.

Then we point Cowork at that folder and ask it a boring question: "Read all of this and tell me how we write proposals." Watching the answer come back is the moment the room changes. It's usually accurate, occasionally embarrassing, and it's the first time most people have seen their own process described from the outside.

The lesson lands before we build anything: the quality of the agent is the quality of the folder. Garbage in, agent out.

Teaching it your way of working

A Project in Cowork is a workspace with its own files, its own instructions, and its own memory. That's where the team's way of doing things lives. Instructions like "we never quote below the grid without a named approver," or "every proposal ends with the three references from the same sector." Written once, in plain language, in French or Darija if that's how the team thinks.

Skills are the next layer, and this is where the training earns its price. A Skill is a repeatable procedure the agent can run on request: "prepare a first draft proposal from this brief," "reconcile this month's supplier invoices against the purchase orders," "turn these meeting notes into the client follow-up in our tone." A good Skill is small and specific. One participant, a communications lead, ended up with three separate writing Skills instead of one big one: the house voice, the formal corporate register, and the newsletter. Each one is a paragraph of instructions and a couple of examples. Each one is used every week now.

Plugins bundle Skills with connectors and sub-agents into something you can hand to a colleague. Scheduled tasks run in the cloud, so the Monday morning "summarise what landed in the shared folder this week" happens whether or not anyone's laptop is open.

None of this is code. That's the point of the tool, and the point of the training. The people who build the best agents in these sessions are not the technical ones. They're the ones who can describe their own job precisely.

The two things I insist on

The first is permissions. Cowork has three modes for actions that touch your files and connectors: it asks you every time, it decides on its own within safe limits, or it skips the check entirely. Everyone wants the third one by the end of day one. Nobody gets it from me until they've watched the agent make a wrong call in the second mode and understood why it happened. An agent with access to your files and your email is not a toy, and security researchers have already written about what happens when connectors are trusted too widely. Grant one folder, not the whole drive. Grant the connector you need, not all of them.

The second is an owner. Every agent we build in a session gets a name on it before the session ends. Not the person who's most excited. The person who'll notice when it drifts, who'll update the examples when the pricing grid changes, who'll be asked "why did it say that" and be able to answer. Without that person, the agent is a demo that stops working in March.

What people come back with

The follow-up call is usually three weeks later, and the pattern is consistent.

The proposal agent works. The invoice reconciliation works. What surprises people is the third thing they built without me: a Skill for a task I'd never have thought of, in a corner of their business I don't know. A logistics manager who taught it to read driver WhatsApp reports and flag the late ones. An HR lead who built a Skill that drafts onboarding checklists from a job description and the last three checklists.

That's when I know the training took. Not when they can use my agent. When they build their own, from their own files, for a problem I'd never have seen.

Your data is already an agent. It's just been sitting in a folder, waiting for someone to ask it the right question. If you want to spend a day finding out what it knows, I run that day.