An agent in a folder

In our guides on Cowork, skills and connectors. Cowork gives the AI a place to work with files, skills capture your way of working, and connectors provide access to more context.

One concept keeps coming back in our Claude Cowork and Claude Code training: the agent in a folder. It sounds almost banal, and it changes what the AI can do for you.

Most people use folders as storage. With tools like Cowork and Claude Code, the folder becomes something more. It becomes the place where the AI finds its sources, follows the rules, builds on previous work and leaves better output behind for next time.

What does agent in a folder mean?

An agent in a folder is simply an AI with a bounded workspace, clear instructions and the right context. It sees the files that live there. It can read the brief, use the examples, follow the instructions, create new files and build on what already exists. Over time the folder becomes more than a collection of documents. It becomes a way of packaging work.

Regular chat starts with you dragging context into the conversation: pasting notes, explaining background, uploading files. With an agent in a folder you move the context out of the chat and into a workspace the AI can use directly.

The effect compounds. Every time you work in the folder, better instructions, better examples, better checklists and finished deliverables can be saved back. Next time the AI does not start from zero. It starts from a slightly better version of the workspace.

It is still about context

Context is the word we say most in our training, because the right context is still the biggest factor behind good results with AI. Many try to solve quality with better prompts. That helps, but only so far. If the AI lacks sources, examples, rules and previous work, the prompt has to carry too much.

A good folder does the opposite. It lets you make the prompt shorter because the context already sits in the right place. You point the AI at the folder and say what you want out: "Write a client summary." The prompt stays short because the folder carries the rest.

The folder gives the AI three things a prompt alone cannot:

• Living context. The context updates after every interaction, as feedback and finished deliverables are saved back into the files.
• A clear workspace. The AI knows what to work with, which reduces noise and makes it easier to control what gets used.
• Finished deliverables. Output lands as files, ready to use.

You do not need to design the structure

The point is not to design the perfect folder structure yourself. That is exactly what Cowork can help with. Explain what you want help with, which files the agent can use, and what you want as the result. For example:

I want an agent that helps me with client follow-up in this folder. Help me set up a good structure with folders, instructions, examples and output.

Claude will suggest a scaffold: a sources folder, an instructions folder with a working method and quality checklist, an examples folder, and an output folder for the deliverables. The structure follows the task, not the other way around. You adjust it through actual use.

Give the agent memory

If you do only one thing, ask Cowork to create a CLAUDE.md at the root of the folder. It is the instruction file Claude reads to understand how to work in exactly this workspace. Think of it as a short onboarding to the folder.

CLAUDE.md can be updated as you work. When Claude gets something wrong, the feedback goes into the file. Over time it also becomes a map of the other files in the folder, so the AI finds the right context faster. In other agentic tools the same file is usually called AGENTS.md.

You take what you would otherwise have explained in the chat, and put it where the AI actually works.

Feedback becomes part of the system

The most important thing happens after the first delivery. If the AI missed something, the feedback should not stay in the chat. Save it in the folder. It can be as simple as saying "update CLAUDE.md with what we learned just now". That is when the folder starts gaining value over time.

From conversation to system

In chat, you explain the task again every time. With an agent in a folder, you build a workspace that improves its own context over time. The AI still will not be flawless; you still review quality, give feedback and own the result. The difference is that the feedback becomes part of the system instead of disappearing into a conversation.

For simple things it is overkill. For complex, repeated work, it is the setup that pays back.

First published in Norwegian on LinkedIn.