Claude Code is only as good as the files you give it. Here is the exact folder system we use at Business Builders so every client engagement is readable by our team and our AI, from kickoff to reporting.
When a teammate or a Claude session opens a client folder, it should already know where everything lives. No hunting. No tribal knowledge. Two decisions make that possible.
Same folders, same names, same places. The value comes from consistency. One weird client folder breaks the pattern for the team and the AI.
Markdown is what an LLM reads best. Facts live in .md files. Long prose can live in Google Docs, but the substrate the AI works from stays markdown.
Everything lives under one root folder, organized by company rather than by file type. A CLAUDE.md file at the root explains the whole structure to Claude: what each folder is, how projects are named, where new work goes.
Claude reads it automatically at the start of every session. It is the employee handbook the AI actually reads.
Every client gets the same folder skeleton. The CLAUDE.md inside records what Claude needs to do good work: tier, retainer, current focus, who is on the account, and any hard rules.
Start a session in that folder and Claude is briefed before you type a word. A reusable skill scaffolds the whole thing from a template with placeholder tokens, so a new client folder takes one command.
The team-facing version lives in a Google Shared Drive, one folder per client, mounted locally through Drive for Desktop. Claude writes files with normal file tools and Drive syncs them up. No API work needed.
The drive also holds a blank template, a fully filled example client, a registry indexing every client, and an AGENTS.md telling any LLM or teammate how to use the library.
Every fact lives in one file. Edit it there, nowhere else. Duplicated facts drift, and drifted facts end up in client deliverables.
Never convert the markdown to Docs. Google Docs is fine for long prose, but the files the AI reasons from stay plain markdown.
Any stat that has not been checked gets a "(verify before publishing)" tag, so it never ships to a client by accident.
Every file carries a small header with owner, last updated date, and status. Stale files announce themselves.
Any team member, or any Claude session, can pick up any client cold. The folder is the briefing.
Onboarding a new client takes minutes. The scaffold skill builds the documented folder in one command.
Deliverables get better. Claude always has the brandscript, voice rules, and strategy in reach.
The system compounds. Every report, play, and decision gets filed where the next session will find it.
Then write a CLAUDE.md at the root that explains the structure, the naming rules, and where new work belongs.
Include a CLAUDE.md inside it with placeholders for tier, retainer, team, and current focus.
Setup becomes one command: copy the template, answer a few onboarding questions, and every placeholder gets filled.
The tooling is simple. The discipline is the product.
Agency Builders is the private community where agency owners trade real systems like this one, operator to operator.
Join the Community