CLAUDE.md: the orchestrator

CLAUDE.md is a plain markdown file in your project folder. Claude Code reads it automatically at the start of every session, before you type anything.

That is the whole mechanism. There is no config, no syntax, no schema. It is a note to your colleague, and your colleague actually reads it.

The mistake almost everyone makes

The first version of a CLAUDE.md is usually a dump. Everything the person knows, in one file, in no order. It grows to 400 lines and stops working.

It stops working for a simple reason. A long file competes with itself. Rule 12 and rule 88 disagree, and neither of you noticed, because nobody reads a 400-line file top to bottom.

The five sections that work

This structure comes from a file that has survived a hundred-plus articles and ten clients. Copy the shape, not the content.

# CLAUDE.md – <project> article system

## 1. Intake (never skip, never guess)
   The questions to ask before any work starts.

## 2. Pipeline (the order is fixed)
   The stages, in order, and what each hands to the next.

## 3. Non-negotiable principles
   Rules that apply to everything. Short. No exceptions listed here.

## 4. Every article ships with
   The checklist. Tables, FAQs, link counts, takeaways.

## 5. File map
   Where everything else lives, one line each.

Make it ask, not assume

The single highest-value thing in the file is the intake section. Not because the questions are clever, but because it stops the model guessing.

A guessed client means a wrong product description. A guessed reader means the wrong reading level. Both are invisible until someone who knows the client reads the draft.