Run claude doctor after installing, stay in ask-first mode, and paste a short CLAUDE.md into one project folder. That is the whole safe setup, it takes about an hour, and at the end Claude Code can describe your files in plain words without changing a thing.
Do the plan decision first. Claude Code needs a paid Claude plan: Pro (about $20 a month) or API access with a card. The free chat plan will not work (Gradually). Start on Pro. Move up when you hit a usage limit in the middle of a task two weeks in a row, not before.
Install it with one command, then run claude doctor
On Mac or Linux, paste this into a terminal:
curl -fsSL https://claude.ai/install.sh | bash -s latest
On Windows, paste this into PowerShell:
irm https://claude.ai/install.ps1 | iex
The install takes about two minutes. Make a new folder for your first project, open the terminal inside it and type claude. A browser window opens so you can log in with your Anthropic account. Back in the terminal, run claude doctor. If it reports no problems, you are set (Go9x). Installer lines change, so if the command fails, copy the current one from Anthropic's Claude Code docs.
Always launch it inside one project folder, never your home folder or Desktop. It can see and change files wherever you start it, so a small folder limits the damage.
Copy the folder first, because a copy is your only real undo
If you cannot read a diff, you cannot judge an edit, and that is most non-developers. Before any big change, duplicate the project folder, or use version control if you already do. Do this before you touch the permission settings below.
Never use bypass mode, and approve 10 prompts in a row before you loosen anything
Claude Code has permission modes that decide when it stops to ask you. Here is the ladder, from the first week up:
| Mode | What it does | Use it when | What it costs you |
|---|---|---|---|
| Default (manual) | Asks before writing files, running commands or making web requests. Reading and searching need no prompt. | Your first week. | Clicking yes without reading wipes out the protection. |
| Plan | Shows a plan and runs nothing until you approve. It can read and search but cannot write or run commands. | The task is big or unclear. | One extra read of the plan before each change. |
| Accept edits | Approves file edits on its own. Commands still ask. | Small changes you can check in your browser. | Edits pile up unreviewed, so you cannot tell which one broke things. |
| Bypass permissions | Removes every safeguard. | Only in a sandbox or container with nothing valuable in it. | A destructive command runs with no question. |
The first three rows follow Anthropic's permission modes page, which also lists the current key for cycling modes mid-session. The bypass warning follows Morph's write-up, which shows a command that can wipe a home folder and notes that nothing blocks instructions hidden in text Claude reads.
Skip bypass mode on any computer that holds your real files. Stay in default mode until you have approved 10 prompts in a row and understood each one. Then try accept edits for page changes you can check in a browser.
Ask three questions before you approve any prompt
- Can I undo it?
- Does it touch anything outside this project folder?
- Does it delete or send anything?
If the answers are no, yes, yes, type "Explain this in plain words first" and wait. The 10-prompt switch point and these three questions are my own rules of thumb. The setup steps and the modes come from the docs and guides linked above.
Paste this CLAUDE.md before your first task
CLAUDE.md is a plain text file in your project folder. Claude Code reads it at the start of every session, so it works as memory. A file in your root folder holds who you are across all projects, and one inside each project folder holds that project's details (Gradually).
Create a file named CLAUDE.md in the project folder and paste this in. Change the bracketed parts.
# About this project
[Two sentences: what you are building and who it is for.]
# About me
I am not a developer. Explain things in plain words and tell me what each step does before you do it.
# Rules
- Ask before deleting or renaming anything.
- Before you change a file, list the files you plan to touch.
- Keep each change small. One thing at a time.
- If you are unsure, ask me instead of guessing.
# How to check the work
[The one page or command I open to see if it works, for example: open index.html in my browser.]
Keep it short. Put in only what you would otherwise retype every session. Skip it and you repeat your context each time and get different answers from one day to the next.
Type these three prompts in your first session
- "Look at this folder and explain what is in it, in plain words. Do not change anything."
- "Before you change anything, list the files you plan to touch and wait for my yes."
- "Do not delete anything. If something should be removed, tell me and I will decide."
Prompt 1 is your first task. It costs nothing, proves the setup works, and shows you how Claude Code describes your files.
Spend your first hour like this
This is a suggested split to try, not a measured result.
| Minutes | Do this |
|---|---|
| 5 | Pick a plan and install it |
| 5 | Run claude doctor and log in |
| 10 | Write CLAUDE.md |
| 10 | Run prompt 1 and read the answer |
| 30 | One small real task in plan mode, then default mode |
A small real task is something you can check in your browser in five minutes: fix a typo on a page, rename a heading. Pick something that small and finish it before you plan an app.
Five mistakes that cost the most, ranked
| Mistake | What it costs |
|---|---|
| Running it from your home folder or Desktop | It can touch far more than the project, and you may not notice |
| Using bypass mode on your everyday computer | A destructive command runs with no question |
| Clicking yes on every prompt | You lose the one safeguard ask-first mode gives you |
| Starting a whole app on day one | A stalled project and a long cleanup |
| Paying for a bigger plan before you know your usage | Money for headroom you may never touch |
If you want to see what a team of agents looks like once the basics feel boring, I wrote that up in How I Run a Portfolio of Small Apps With a Team of AI Agents.
Do this in the next ten minutes
Make a new empty folder, open a terminal inside it, install Claude Code with the command above and run claude doctor. Then type prompt 1 and read the answer. You do not need a real project yet. You need to see it run once.
Got a job you wish a tool would do for you?
If a manual task is eating your hours and you are not sure it can be automated, send me a short note describing it. Email Me