Skip to main content

Claude Code Basics Lesson 26 of 46

How to Write a CLAUDE.md That Actually Works

In short
  • Short beats thorough: past a certain length, instructions get followed worse.
  • Write things that are verifiable, not wishes: "files go in X" instead of "keep things tidy".
  • You add a line on the second repeated correction, not by anticipating.
  • Real maintenance is removing, not adding.

The trap: writing everything into it

The moment you understand what it's for, the temptation is irresistible: write everything into it. How you work, what you like, every exception, every edge case, the project's history.

And here's the counterintuitive part, the one worth the whole lesson: a longer file doesn't get followed better. It gets followed worse.

Two reasons, and you already know both. First: that file enters the working space at every session, so the longer it is, the more room it takes. Second, and more important: buried among two hundred lines, the three that actually matter carry less weight. The signal dilutes into the noise — exactly like in a conversation that's run too long.

The practical benchmark: under two hundred lines. If you're going over, you don't need more room: you need to cut.

Write things that are verifiable, not wishes

The difference between an instruction that works and one that's almost always ignored comes down to this: can you check whether it was followed?

  • "Keep the files organized" → "New contracts go in Documents/Contracts/year, named client-date"

  • "Write professionally" → "In emails: no exclamation marks, always close with name and role"

  • "Be careful with numbers" → "Every total gets double-checked by adding it up again before writing it down"

The first ones are wishes: there's no way to know if they were honored, so there's no way to honor them. The second ones are checks.

What goes in and what doesn't

  • Goes in: how your stuff is organized, naming conventions, things to always do, the commands you use, what things are called in your world.

  • Doesn't go in: long step-by-step procedures. They take up room in every session even when they're not needed. There's a tool made just for those, and it's the chapter right after this one.

  • Also doesn't go in anything that's obvious from looking at the folder: if it's visible on its own, writing it down is wasted space.

When you add a line

Not by anticipating. The file grows from experience, and the signal is precise:

  • you've corrected the same thing for the second time;

  • you've caught yourself typing in chat a sentence you'd already typed the session before;

  • a new person, to do your job in your place, would have needed to know that exact thing.

If any of these is true, that line earns its space. Otherwise, wait.

Real maintenance is removing

There's a habit that sets apart people who use these tools well: they remove things every once in a while. Try deleting a rule and see if it gets that wrong again. If nothing happens, that line wasn't doing anything for you — and you were paying for it every session.

Two signs it's time to clean up: instructions that contradict each other (if two lines say opposite things, it'll follow one at random) and stale instructions, ones that describe how you worked six months ago.

What's next

I've said twice now that things "take up space", and I've been acting like you already know what space I mean. In the next lesson we actually look at it: the context window, meaning how much it can hold on the table at once.

Did you get it?

Two minutes, no grade. Trying to answer — even getting it wrong — cements the concepts far more than rereading. Not in the mood? Skip it: the lesson still counts as complete.

A very long CLAUDE.md that explains everything: what happens?

Which of these two instructions is well written?

When is it worth adding a line to CLAUDE.md?

Course Claude Code Basics
  1. Chapter 1 Introduction

    1. 1 The Claude Ecosystem and Where Claude Code Fits
    2. 2 What Claude Code Is: The Collaborator Who Walks Into Your Studio
    3. 3 How an Agent Works: The Loop
  2. Chapter 2 What's Underneath

    1. 4 Claude Code in Action: The Loop, for Real
    2. 5 The Model Has No Memory: It Starts Over Every Time
    3. 6 The Harness: The Engine and the Gear Around It
    4. 7 The Map: Four Buckets
  3. Chapter 3 You Choose the Engine

    1. 8 There's No Single Claude: The Model Lineup
    2. 9 Which Model for Which Job
    3. 10 Effort: How Hard It Should Think
    4. 11 Model or Effort? The Question That Trips Everyone Up
    5. 12 The Bill: Why Saving Money Can Cost You More
  4. Chapter 4 First Time at the Controls

    1. 13 Where to Run It: Four Doors In
    2. 14 What You Need Before You Start
    3. 15 Installing Claude Code
    4. 16 The Folder: Where You Open It Decides What It Works On
    5. 17 Choosing the Engine, in Practice
    6. 18 Your First Task: Small, Yours, and It Works
  5. Chapter 5 What You Let It Touch

    1. 19 Why It Asks for Permission
    2. 20 The Rules: Always Yes, Ask Me, Never
    3. 21 The Modes: How Much Leash You Give It
    4. 22 Plan Mode: Get the Plan Before the Work
    5. 23 Verification in the Loop: Give It a Yardstick
    6. 24 Example: Putting the Bucket Together
  6. Chapter 6 What It Knows About You

    1. 25 CLAUDE.md: The Notes It Reads the Moment It Walks In
    2. 26 How to Write a CLAUDE.md That Actually Works
    3. 27 The Context Window: How Much It Can Hold at Once
    4. 28 Cleaning Up: Compact and Clear
    5. 29 Automatic Memory
    6. 30 Example: From Repeating Yourself to Never Again
  7. Chapter 7 Skills

    1. 31 What a Skill Is: A Procedure Written Once
    2. 32 Creating Your First Skill
    3. 33 The Description Decides Whether It Gets Used
    4. 34 Example: A Skill That Saves You the Week
  8. Chapter 8 How Many You Put to Work

    1. 35 The Subagent: A Helper With Its Own Desk
    2. 36 When a Helper Is Worth It (and What It Costs)
    3. 37 Teams: When Agents Talk to Each Other
    4. 38 Example: Searching Without Cluttering the Desk
  9. Chapter 9 What It's Connected To

    1. 39 MCP: The Standard Plug Into Your Tools
    2. 40 Plugins: Packaging Up and Borrowing
    3. 41 Example: When You Stop Copying and Pasting
  10. Chapter 10 Automating the Setup

    1. 42 Hooks: The Thing That Always Happens
    2. 43 Where It Fires and What It Can Do
    3. 44 When You Actually Need One (and When It's Too Much)
  11. Chapter 11 Wrap-Up

    1. 45 The Project: Your Expense Tracker
    2. 46 Wrap-Up: What You're Taking Home
From a Confusing Brief to a Complete UX/UI A vague brief becomes a navigable app. Product Ad From a Single Photo A photo becomes an animated ad, no code required. Get Claude to Watch Your Videos Claude watches your videos and turns them into text. Higgsfield Inside Claude Code Generate images and video while you code in Claude Code. Turn a Loom Recording Into a Web Page A screen recording becomes a web page, no code. Vertical Shorts With NotebookLM Your sources become a vertical short. Luxury Landing Pages on Lovable A luxury landing page from a single prompt. Excalidraw Running Locally Excalidraw free on your computer, no code. Mistral OCR in Your Workflow Extract text from documents with Mistral OCR. Claude SEO in the Terminal 25 free SEO skills inside Claude Code. Google Search Console inside Claude Code Search Console data inside your terminal. Claude Code Routines Claude working on its own, computer off. Clone a Landing Page in React With v0 Clone a real landing page into React code. Context Economy With Claude Work light and don't burn through Claude's limits. From Prompt to Self-Improving Skill Claude skills that learn from your mistakes. From NotebookLM to Canva: Presentations NotebookLM slides, finally editable in Canva. The Map for Understanding Every AI Tool 12 categories for placing any AI tool. Transparent PNGs With ChatGPT Real transparency, not a fake checkerboard. Animated Infographics With Gemini Infographics that loop, animated with Gemini. Market Research With Deep Research Deep Research as your market analyst. Get Cited by AI Search Engines (AEO) Become a source that ChatGPT and Perplexity cite.