AI Automation Academy
Free download
The CLAUDE.md I use
The file that tells Claude Code how you want your project run, with the reasoning left in.
The file is on the next page.
Just heard me mention my CLAUDE.md? This is it. If you
have not run into one before: a CLAUDE.md is a plain text file
you keep in your project folder. Claude Code reads it every time you start
work, so it already knows how you want things done instead of asking you
all over again.
This is step one of every build I walk through
- Make a new folder for whatever you are building.
- Drag this file into it.
- Open that folder in Claude Code and ask it to set the folder up based on the
CLAUDE.md.
That is the whole setup. After that you type what you want in plain English and it goes. You do not need a coding background, and if Claude Code has ever looked like something you have to be a developer to use, this file is most of what makes it stop looking that way.
# Agent Instructions
One git repo. A handful of top-level categories, one per area of
life or work, plus a shared-tools library — which is not an area
but a library: nothing is *about* it, other categories call into
it by path.
...
**Give every rule its reason** in the same sentence. A rule
without a reason gets argued with.
Why this one instead of a blank file
Most CLAUDE.md files are a list of preferences somebody wrote
once. This one has been sharpened over months of real projects, rewritten
every time it turned out to be wrong about something, and every rule in it
says why it exists, which is what stops you arguing with your own
instructions six weeks later.
What you get alongside the rules:
- Why it writes code instead of guessing. AI is probabilistic: ask the same question twice and you can get two different answers. Code is deterministic, so it does the same thing every single time. At 90% accuracy per step a five step chain lands around 59%, which is why anything that has to be right every time gets handed to code, and Claude's job becomes deciding which code to run rather than doing the work by hand.
- Self-healing, not just retrying. When something breaks it does not work around it. It finds the cause, repairs it, and updates its own instructions so the same failure cannot happen twice. It also knows what not to chase, because a network timeout is weather, not a bug.
- How to work with it as a partner, not a spec sheet. Ask it for a recommendation instead of a menu, a cheap version to react to before it builds the real one, and something you can edit rather than a wall of chat.
Already subscribed? Enter your email anyway. You land on the same page with the same file, and you will not be added twice.