A non-engineer built a five-agent AI team using nothing but markdown files and Claude Cowork, a desktop app by Anthropic. No code, no terminal, no configuration screens. The entire system lives in a single 106-line text file called CLAUDE.md stored in a Dropbox folder. Setup cost roughly 90 euros in tokens and took one day. The agents have names and distinct roles: Elke runs editorial, Joan handles sales and pricing, Caitlin manages voice and writing, Miranda owns product and design, and Rachel advocates for the reader. They read files, edit documents, create new ones, and write session notes so context survives between conversations.
What makes this worth reading in full is not the conclusion but the construction. The author details exactly what goes inside CLAUDE.md, including a single instruction that separates a generic chatbot from five agents that push back on decisions. The key line in the philosophy block is: 'Be honest, not helpful.' That instruction alone changes agent behavior in ways the piece walks through specifically. The agents disagree with each other, flag lazy thinking, and update their own files between sessions. The author also addresses the ethics of using real first names as persona anchors without copying any actual person's writing.
The method is not Claude-specific. Any AI tool that reads project files can run this structure. The author has since built additional teams for other work areas and is building one to oversee their entire business. The barrier to entry is a text file and a conversation. The article includes the full file structure, the exact agent descriptions, and a step-by-step starting point. If you work alone and have ever wished for a team, this is the most concrete implementation guide currently available from someone who had no prior engineering background.
[READ ORIGINAL →]