One File That Makes Claude 10x Smarter in Every Cowork Project

Cover image

You open Cowork, you give Claude a task, and it does it like a stranger

Every time you explain the same things again, who you are, what your business does, the voice you write in, the rules you always follow, and the moment the task is done, all of it is gone

Meanwhile there's one small file that could sit in your project and make every task sharper without you saying any of that again, and almost nobody creates it

That file is CLAUDE.md, and it's the cheapest upgrade to your output in the whole tool

This article shows you exactly what it is, what goes in it, and hands you a ready one to drop in today

Part 1: What CLAUDE.md is in Cowork

Forget the developer version for a second, this has nothing to do with a terminal or code

In Cowork, CLAUDE.md is just a plain file that lives in your project, and Claude reads it first thing, every single session, before it touches your task

Think of it as the always-on background: Project Instructions and one-off prompts come and go, but CLAUDE.md is the standing context that's there for every task in that project, so Claude acts like it already knows you

You set it up once, and it works in every task from then on, that's the whole leverage, one file, written a single time, quietly making everything you do in that project better

And it has to be a file, not context you paste into each chat, a pasted block dies the moment the session closes and tomorrow you're explaining again

A file stays put and Claude picks it up on its own every time, the tool does the remembering instead of you

Part 2: The 3 places context lives in Cowork

Here's where most people go wrong, Cowork gives you three places to put context and they dump everything into one, or use none, so Claude keeps acting like it doesn't know them

Project Instructions, the rules and behavior, how to format outputs, what to prioritize, what to skip, this is the how
CLAUDE.md, who you are and how you work, your business, your voice, your workflow, the operational playbook, this is the who and the what
Skills, the task specialists, they fire on their own when a request matches, "turn this into a thread" wakes the thread skill, this is the when

The difference that matters: Skills act only when a task triggers them, CLAUDE.md is on for everything, it's the layer underneath all the others, the foundation the rest stand on

Get this split right and Claude stops feeling generic, the rules live in Instructions, the identity lives in CLAUDE.md, the repeatable tasks live in Skills, and nothing gets re-explained

Part 3: What actually goes in a CLAUDE.md

A CLAUDE.md isn't a wall of text, it's a handful of sections that tell Claude who it's working for

Who you are, your name, your business, what you actually do all day
Who you serve, your audience or your clients and what they want
Your voice and rules, how you write and communicate, the words you use and the ones you ban
Your tools and workflow, what you work with and how a normal project moves
What good looks like, the standard a finished piece has to hit before it's done
What to never do, the mistakes you're tired of correcting

Each line you add is a thing you never have to type into a chat again, that's the trade, five minutes writing it once against a hundred re-explanations later

Part 4: The before and after

Here's the difference in one task, say you ask Claude to write a launch post for your new service

With an empty project, Claude has nothing to stand on, so it gives you the average of the internet, corporate lines, hedged phrasing, words like unlock and game changer, a post that could belong to anyone and lands with no one

With a CLAUDE.md in the project, Claude already knows you're a freelancer who writes in short punchy lines, always through a money angle, never with filler, and it knows the exact words you refuse to use

Same request, same model, but the second post comes back in your voice, on your angle, ready to publish instead of ready to rewrite

You didn't prompt better, you set the ground once and every task after it inherited that ground, that's what the file buys you

Part 5: How to set it up in Cowork

Step 1: open the project you work in, or make one for the client or the kind of work you do most
Step 2: create a file named CLAUDE.md in the folder your project is connected to
Step 3: fill in the sections from Part 3, keep each answer short and concrete, real specifics beat vague description every time
Step 4: run one normal task and see how it comes back, notice where Claude still missed you
Step 5: add the missing detail to the file, that's the whole maintenance, every gap you patch makes every future task better

Two or three rounds of this and the project knows you well enough that you stop noticing you set it up, the work just starts coming back right

And keep feeding it, every time you catch yourself fixing the same thing in an answer, add that fix to the file as a rule, the CLAUDE.md gets sharper the longer you use it, it's the one file that quietly compounds

Part 6: A ready CLAUDE.md to copy

Drop this into your project, swap everything in brackets for your own answers, and you've got a working setup in five minutes:

# CLAUDE.md

## Who I am
[your name, your business, and what you do]

## Who I serve
[your audience or clients, and what they want from you]

## My voice and rules
- [how you write, for example short punchy lines, no filler]
- [your angle, for example always practical or money-focused]
- [words or styles you ban]
- [format you prefer]

## My tools and workflow
[the tools you use and how a normal project moves from start to finish]

## What good looks like
[the standard a finished piece has to hit before it counts as done]

## Never do
- [a mistake you're tired of correcting]
- [another one]

Change the brackets to your own answers, save it in the project, and every task in that project starts from this instead of from zero

Part 7: What a filled-in CLAUDE.md looks like

The blank is easy to skim past, so here's a real one, a freelance web designer, copy the shape and swap in your own answers:

# CLAUDE.md

## Who I am
I'm a freelance web designer, I build websites and landing pages for small businesses and founders

## Who I serve
Founders and small business owners who need a clean site or landing page fast, they care about it converting, not about design theory

## My voice and rules
- clear and direct, explain choices in plain language, no design jargon
- always tie a design decision to the client's goal, usually leads or sales
- keep copy short and benefit-first
- clean modern layouts with lots of whitespace, no clutter
- mobile first, most of their visitors are on a phone

## My tools and workflow
I design and prototype in Cowork, a project goes brief, wireframe, copy, build, then a review pass before it goes to the client

## What good looks like
A page the client can launch as is, loads fast, one clear call to action, reads well on a phone

## Never do
- add sections or features the client didn't ask for
- use stock corporate copy or filler text
- ship a design without checking it on mobile

Notice it's short and specific, that's the whole point, six tight lines Claude can act on beat a page of vague description every time

Part 8: The mistakes that make a CLAUDE.md useless

A bad file does nothing, or worse, it confuses Claude, here's what kills it:

Too long, a wall of text and Claude drowns, keep it to a page at most, every line earns its place
Too vague, "write well" tells Claude nothing, "short sentences, no corporate words, under 150 words" it can actually follow
Contradicting your Instructions, if the file says one thing and Project Instructions say another, you get mush, keep each rule in one place
Stuffing everything in, it's not a diary, it's a playbook, only what changes how Claude works for you belongs in it
Writing it once and forgetting it, a file you never update goes stale, the good ones grow every week

Fix these and the file earns its keep, break them and you've just added noise Claude has to read before it even starts

Conclusion

Many people treat Claude like it has amnesia because they never gave it a memory, they re-explain themselves every session and blame the tool for sounding generic

The fix isn't a better prompt, it's one file, written once, that tells Claude who it's working for and stays on for everything you do in that project

Just select the project you work in most, spend 5 minutes on the template above, and watch the next task come back already knowing you

That's the smallest change in Cowork with the biggest difference in what you get back