1. What are these files?

The heart of the system is a set of .md files (Markdown — plain text with light formatting) that you paste into the "Project Instructions" of a Claude project. The moment you do that, every conversation in that project starts with this context.

Claude knows your style, your industry, your ground rules. You no longer have to re-explain who you are every time. That's the difference between a generic assistant and one that actually knows you.

2. The principle: Compound Engineering

The idea comes from a concept called Compound Engineering: every time Claude does something that annoys you or gets your style wrong, you add a rule. The file improves with use, and all your future exchanges benefit from it.

That's the real lever: you're not "prompting" better one time — you're building an asset that compounds over time. After a few weeks, Claude nails on the first try what used to take you three back-and-forths.

In one sentence

A good Claude setup isn't a better prompt — it's a system of files that learns your preferences and applies them automatically to every conversation.

3. How it actually works

There's a main file — CLAUDE.md — that defines the essentials: your tone, your response format, your behavior rules, your professional context. It's the base that runs all the time.

Then there are extension files, loaded only when they're useful:

CLAUDE-strat.md — for business / comms / marketing topics.
CLAUDE-dev.md — for anything code-related.

The main CLAUDE.md knows when to call one or the other — you don't load everything every time. That's what keeps the context light and relevant.

I also have a file dedicated to the brand's story and DNA. That's what Claude loads when it needs to write, pitch, communicate — it knows the positioning, the tone, what we say and what we don't. It's 1 hour to write with Claude and it radically changes the quality of everything it produces about your brand.

Hugo Lahutte at the mic, presenting how to use Claude to merchants — On stage: showing, concretely, how to work with Claude.

4. Claude Chat vs Claude Code

This file system is for Claude Chat — the web interface at claude.ai. You create a Project, paste your instructions, and you're off.

Claude Code is something else: a command-line tool that runs in your terminal, right inside your repo. It sees your files, understands your codebase, and can run code. If you do dev work, it's a level up. The config principle is the same, but the use is 100% technical.

5. Where to start

Look at existing files for inspiration, but don't copy them as-is — they're tuned to a context that isn't yours. My recommendation:

Create a Project on claude.ai.
Ask Claude to build your CLAUDE.md by pasting the prompt below.
Once the foundation is solid, tackle the brand-DNA file — story, positioning, tone. That's where it gets really powerful for comms and content.

The copy-paste prompt to generate your file:

Prompt — generate your CLAUDE.mdI want to create my CLAUDE.md file — a personal configuration file for my Claude projects. I'm attaching reference files CLAUDE.md, CLAUDE-strat.md and CLAUDE-dev.md. Read them to understand the structure, the level of detail expected and the overall logic — but don't copy them: they're tuned to a different context, not mine. What I want to build: 1. A main CLAUDE.md that covers: - My preferred tone and format (formal or casual? prose or bullets? how long should answers be?) - My professional context: what I do, my company, my industry, my role - The behavior rules to enforce (web search, honesty, no flattery, push back on bad ideas) - What I do NOT want (preambles, end-of-answer summaries, unrequested frameworks, over-complication) - The logic for calling the extension files: this file must know when to load the others depending on context 2. A [MyCompanyName].md file describing my company's DNA: - Story and context - What we do, for whom, why - Positioning and values - Brand tone: what we say, what we don't - This file will be loaded whenever Claude produces content related to my brand 3. If I have recurring technical or domain needs, we'll create the appropriate extensions (the equivalent of CLAUDE-dev.md or CLAUDE-strat.md) once the foundation is in place. Method: ask me ONE question at a time. We build section by section. When we've covered everything, you generate each complete file in markdown, separately.

That's all you need to get started. The rest builds itself through use: every friction becomes a rule, and your system gets better on its own.

How I use Claude