3 Claude Code Habits for Better Work
Claude grades its own homework, makes the same slow mistakes every time, and forgets you the minute you close it. These 3 habits fix all of that with nothing to install: the exact prompts, the brain file setup, and a walkthrough.
Most people try to get better work out of Claude by adding more. More plugins, more skills, longer prompts.
That doesn't fix the real problems. Claude checks its own work, so it misses its own mistakes. It makes the same slow mistakes every time. And it forgets who you are the minute you close it.
How you run Claude matters more than what you install. These are the 3 habits, with the exact prompts. You paste a prompt and answer a few questions.
Step 1. Make one folder for your business
Make a new folder on your computer called my-business.
Open the Claude desktop app, switch from Chat to Code at the top, and pick that folder.
Everything below happens in that folder. Claude Code reads the files in it every time you start. When Claude asks to create or edit a file in that folder, say yes.
Step 2. Set up your brain file
Nothing you do in Claude matters if you have to re-explain yourself every single time.
So you give it one file that says who you are and how you work. It's called CLAUDE.md, and Claude Code reads it every time you open that folder.
This is the first half of habit 3. Do it first, because the other two habits use it.
Paste this. Claude interviews you in 2 short rounds, then writes the file. Short answers are fine.
Build my Claude Code brain file. It's one file called CLAUDE.md in this folder. It says who I am and how I work, so I never have to explain myself again. If CLAUDE.md already exists, read it first, keep what's in it, and only ask me about what's missing. Interview me first, in 2 short rounds. Ask one round at a time, and wait for my answers. Round 1, who I am: my name, what I do and who I help, what I sell and the prices, my real results (only ones that really happened), and the facts you must never get wrong. Round 2, how I work: how I like your answers, words or things I can't stand, what you must always check with me first, and the tools and files I use. Then write CLAUDE.md in this order. The first line is exactly @HANDOFF.md (it makes every new session read my handoff note first). Then these parts: ## Who I am ## Facts I never want wrong ## How I work ## Always check with me first ## Rules I've learned (leave this one empty for now) ## How we work together Put these 4 lines under it, word for word: - At the start of every session, read HANDOFF.md first. If I ask where we left off or what's next, answer from it. If there's no HANDOFF.md yet, skip it and don't mention it. - When I say "save that as a rule", add it under "Rules I've learned" as one short line that says what to do. If a rule there already covers it, update that rule instead. Show me the line. - When I say "wrap up", rewrite HANDOFF.md: what we did, where things stand, what's next (the very next step first), and what I decided. Carry over anything from the old note that isn't done yet. Under 30 lines. - Check every number, price, date, name, result and quote against "Facts I never want wrong". If it isn't there, ask me. Never make it up. Rules for writing the file: - Use only what I tell you. If something is missing, write [ask me]. Never invent numbers, results, stories or quotes. - Keep it short and plain, under 60 lines. When you're done, show me the file.
Here's what your brain file looks like when it's done. This is the same setup I use: one file that says who I am and how I work, rules Claude saved from past jobs, and a handoff note the next session reads first. If you'd rather write it yourself, copy this, fill in the brackets, and save it as CLAUDE.md in your folder.
@HANDOFF.md ## Who I am [your name]. I help [who you help] with [what you help them do]. ## Facts I never want wrong - [what you sell]: [price] - [a real result, only one that really happened] - [a number, date or name Claude must always get right] ## How I work - [how you like answers: short or long, bullets or full sentences] - [words or things you can't stand] - [the tools and files you use] ## Always check with me first - [things Claude must ask before doing, like naming a client or quoting someone] ## Rules I've learned ## How we work together - At the start of every session, read HANDOFF.md first. If I ask where we left off or what's next, answer from it. If there's no HANDOFF.md yet, skip it and don't mention it. - When I say "save that as a rule", add it under "Rules I've learned" as one short line that says what to do. If a rule there already covers it, update that rule instead. Show me the line. - When I say "wrap up", rewrite HANDOFF.md: what we did, where things stand, what's next (the very next step first), and what I decided. Carry over anything from the old note that isn't done yet. Under 30 lines. - Check every number, price, date, name, result and quote against "Facts I never want wrong". If it isn't there, ask me. Never make it up.
Step 3. Habit 1: make Claude check its own work
When Claude writes something and then checks it, it's grading its own homework. It misses its own mistakes.
So you spin up two agents, a writer and a checker. An agent is a helper Claude Code starts for one part of the job. The checker gets fresh eyes. It never saw the writer's thinking, so it reads the draft like a stranger and checks every fact against your brain file.
The writer can't hand it to you until the checker signs off.
That's how my scripts get made. Claude writes, and a checker tests every fact about me against my files before I see it.
Use it for anything people will see: emails, posts, sales pages, captions. Fill in the 2 brackets, then paste:
Spin up two agents, a writer and a checker. The writer can't hand it to me until the checker signs off. The job: [what you want, like "a launch email for my course" or "3 Instagram captions about my new offer"] Files to use: [file names in this folder, or "none"] How it works: 1. The writer writes the draft. If I gave you my own draft or old copy to start from, the checker checks that first, before the writer touches it, so I can see what was wrong. 2. The checker is a separate agent with fresh eyes. It checks every number, price, date, name, result and quote against CLAUDE.md and the files I named. My facts in CLAUDE.md win over anything else, including old drafts. Anything it can't find there is an error, even if it sounds right. It also checks the draft does the job I asked for. It answers APPROVED, or REJECTED with a list of what's wrong. 3. If it's rejected, the writer fixes only what's on the list, and the checker checks again. Up to 3 rounds. If it still isn't approved, show me the draft and the checker's list, and tell me it didn't pass. 4. If a fact is missing and only I would know it, don't guess. Put [ask me: what's missing] in the draft. When the checker signs off, show me the final draft. Under it, list what the checker caught and how it was fixed.
Already wrote it yourself, or reusing an old email? Paste your draft after "The job:", or put the file name in "Files to use". The checker checks yours first.
You get the final draft and a short list of what the checker caught. Read that list. It shows you what would have gone out wrong.
Step 4. Habit 2: make it learn from every job
There are two pieces everybody misses.
First, when a job is done, ask it this:
How could you have done this faster, with fewer tokens? Be specific to the job we just did. Name what wasted time or tokens, and what to do instead next time. Use plain words.
Tokens are the bits of text Claude reads and writes. Fewer tokens means faster work, and your plan's limit lasts longer.
Then make it keep the answer:
Save that as a rule, so you do it that way next time. Add it to CLAUDE.md under "Rules I've learned", as one short line that says what to do. If a rule there already covers it, update that rule instead of adding a new one. Show me the line.
Once your brain file is set up, "save that as a rule" on its own does the same thing.
Now it gets a little better every time you use it.
Step 5. Habit 3: end every session with a handoff
The brain file says who you are. The handoff says where you left off.
At the end of every session, before you close Claude, paste this:
Before we stop, write down what we did and what's next. Rewrite HANDOFF.md (don't just add to the bottom) with: - What we did today, in short bullets - Where things stand: what's done, what's half done, and which files - What's next, the very next step first - What I decided, in my words Carry over anything from the old HANDOFF.md that isn't done yet. Only write what really happened. Keep it under 30 lines. Then make sure the first line of CLAUDE.md is @HANDOFF.md, so the next session reads it first. If there's no CLAUDE.md, make one with just that line.
Once your brain file is set up, you can just type "wrap up" instead.
Next time, open Claude Code in the same folder and ask:
Where did we leave off?
It already read the note, so it tells you what you did and what's next. You don't explain a thing.
Your day with the 3 habits
- Open Claude Code in your my-business folder. Ask "Where did we leave off?"
- Anything people will see goes through the writer and the checker.
- When a job is done, ask how it could have been done faster, with fewer tokens. Then say "save that as a rule".
- Before you close, type "wrap up".
The limitations you should know
- The checker can only check against what's in your files. If a fact in CLAUDE.md is wrong, it checks against a wrong fact. Keep "Facts I never want wrong" up to date.
- A writer and a checker take longer and use more of your plan than one quick answer. Save it for work people will see.
- Rules pile up. Once a month, open CLAUDE.md, read "Rules I've learned", and delete any you don't want.
- Claude Code only reads the CLAUDE.md in the folder you open it in. Open a different folder and it won't know you.
- If you close without the handoff, the next session only knows what's in CLAUDE.md.
Want the next guide? Follow @iamzachcrawford