24 Feb 2026 10 min read LegalTech

Two Files, One Workflow. No Code. Just Cowork.

I thought Cowork was a poor man's Claude Code. Building CoDraft changed that. Two text files in a folder — and when you open it in Cowork, you get a complete workflow. Here's how to build your own.

Photo by Brad Switzer / Unsplash

I'll be honest: I thought Cowork was a poor man's Claude Code. A simplified version for non-coders. You get the comfortable chat interface of Claude or ChatGPT, but now you can install plugins and work on a bunch of documents. I showed in the last post that you can run Cowork with Anthropic's legal plugin. You get NDA review and canned responses. Useful, but limited.

Read the previous post to get started on Claude Cowork.

Then I built CoDraft.

CoDraft is a contract generator tool I've been working on. The inspiration was docassemble — a battle-tested open-source system for guided interviews and document assembly. It's a proper web application with Python, logic blocks, and structured interview flows. I'd used it before and thought: could I get Cowork to do even part of that — find a template, ask some questions, fill it in?

Version 1 is primarily two text files in a folder. That's the whole thing. No Python. No framework. Two text files. But when someone opens that folder in Cowork, they get a working workflow: ask Cowork to prepare an NDA, and it goes searching for an NDA template in your folder, and asks you questions in order to fill it. The best part: I created v1 using Cowork itself. Yes, that "comfortable" chat interface.

I hadn't been wrong about Cowork. The rules had changed, and I hadn't noticed.

In this post, I'll show you how it works — and how you can build your own workflow using nothing but Cowork's chat.

What's inside CoDraft v1

CoDraft v1 is at its core two files. That's it.

codraft/
├── CLAUDE.md                    # Project instructions for Claude
├── .claude/
│   └── skills/
│       └── codraft/
│           └── SKILL.md         # The skill

`CLAUDE.md` — the project instructions

This is a markdown file that sits in the root of the folder. When you open a folder in Cowork, it checks for a CLAUDE.md in the root and loads its contents into Claude's context before your first message. You don't configure this. You don't trigger it. You just open the folder, and Claude is already briefed.

Think of it like the first page of a matter folder — the one that tells a new team member: here's what this project is about, here's how we do things, here's what to watch out for. Except this briefing is for Claude, and it reads it every single time, without fail.

You can read it yourself. It's in English.

A skill — the repeatable procedure

A skill is a markdown file that lives in a .claude/skills/ subfolder. It defines a specific workflow Claude can run — like a standard operating procedure for a recurring task.

If CLAUDE.md is the briefing note, a skill is the checklist. It tells Claude: when someone asks for this, here's exactly how to do it, step by step.

That's the entire architecture. A briefing note and a checklist. Two text files that turn a folder into a workflow.

The difference between project instructions and a skill

If they are all text files and meant to be read by Claude Cowork, what's the point in separating them and keeping the skill in a strange place like under a .claude folder?

It comes down to when they're read. Project instructions (CLAUDE.md) are always read — every session, automatically. Skills are only read when they're requested. You might only need to review a contract when someone asks to review one, so the skill should only be loaded then. But details like where your company is based or instructions not to give legal advice — those apply all the time. They belong in your project instructions.

The folder is the program

This is what I missed when I dismissed Cowork as a "poor man's Claude Code."

In Claude Code, you write code that runs. A proper software project needs infrastructure around it: configuration files, directories for organising code, systems for tracking every change. That's unavoidable overhead when you're actually building software. Claude Code reads CLAUDE.md to understand your project, so it can act as your coding assistant in that environment.

However, if you think about it, none of the above is essential for Claude to be useful. What you really need is a way to tell Claude what to do and for Claude to remember how to do it. That's essentially your folder.

I called it a poor man's Claude Code. That was wrong. It's a stripped-down model. Claude Code for people who don't need to build software — just direct it. Claude Cowork is Zen.

Building it yourself: step by step

Let's create a legal chatbot that summarises contracts and does a preliminary flag of certain risks before legal review. You'll create a folder with a CLAUDE.md and one skill. By the end, you'll have a working Cowork workflow you can share with anyone on your team.

Time required: About 30 minutes (budget more if you want to iterate on the instructions).

What you'll need: