Field guide
Build it yourself
Written for people who have never opened a terminal. Part one doesn't need one. If you haven't read Help it help you, start there.
Before you start
- A paid Claude plan. Pro is $20 a month. The desktop features below aren't on the free plan. Current prices.
- A Mac or Windows computer with the Claude desktop app. Download it here.
- One real project. A newsletter, a monthly report, a client file. Something you do every week or month works best, because the notes pay off on the second run.
- An hour or two to set it up. The habits take a few weeks to stick.
Before you put any business data in, read the privacy section of Help it help you and check your own settings.
Where to start
Start with part one. Claude's desktop app can now work directly with files in a folder you choose, which covers all three habits without a terminal. Move to part two, Claude Code, when you want rules for specific kinds of files, procedures it follows on its own, and the ability to block an action in code.
Part one
Claude's desktop app, no terminal
The desktop app has a mode that works with the files in a folder you connect. Anthropic introduced it as Cowork and has been folding it into the regular app, so depending on your version it may be its own tab or just part of Claude. Either way, it can only read and write in folders you've connected, and it asks before permanently deleting anything. The work itself runs on Anthropic's servers and reaches your folder through the desktop app, so keep the app open while it works.
Step 1 · Write it down
Make a folder for one project
Name it after the project. Everything about it lives here: the notes, the files you work from, and what it produces. Then connect the folder in the desktop app.
Step 2 · Write it down
Have it start a notes file
Type this in your first session:
Create a file called notes.md in this folder. Put in what this project is, who it's for, and how I like things written. Ask me anything you need to fill it in.Step 3 · Write it down
Open and close every session the same way
Start with:
Read notes.md first. Then [what you want to do today].And before you stop:
Before we stop, how do I get you back up to speed as fast as possible next time? Put the answer in notes.md.Step 4 · Give it access
Give it real data, the easy way first
Drop an export into the folder: a spreadsheet from your books, a list of last month's sales, a contract. That alone gets you past copy and paste.
Then connect one tool. Open the + menu in a chat, choose Connectors, pick one, and sign in. Try a question that needs it. When there's no connector for what you need, ask:
How can we get you this data automatically?Step 5 · Ask for a blind review
Check it before it ships
Open a new chat, attach the finished piece and its sources, and type:
Check every claim in this draft against the attached sources. List the problems in order of importance. Don't fix anything yet.Fix what matters, then run it again until it only finds nitpicks.
Keep it safe
Keep sensitive files out of the folder you connect. Anthropic's own safety page says the chance of an attack hidden in a file or web page is "still non-zero," and it advises against using this mode for sensitive information like financial documents. Leave it on the setting where it asks before acting, and read what it's asking. See Use Claude Cowork safely for current details.
Part two
Claude Code, when you want more control
Claude Code is the same idea with more control. Instructions load automatically at the start of every session, you can add standing rules and saved procedures, and you can enforce a rule in code. It runs in a terminal or in the Code tab of the desktop app, and it also needs a paid plan.
To install it, open Terminal on a Mac and paste the first line, or PowerShell on Windows and paste the second. When it finishes, open a new window and type claude --version to check it worked. If you've never used a terminal, Anthropic has a terminal guide for new users.
# Mac
curl -fsSL https://claude.ai/install.sh | bash
# Windows (PowerShell)
irm https://claude.ai/install.ps1 | iexStep 1 · Write it down
Create a project folder and open Claude Code in it
mkdir my-ai-workspace
cd my-ai-workspace
claudeStep 2 · Write it down
Write a CLAUDE.md file
Claude Code reads this file at the start of every conversation, so you don't have to ask. Tell it who you are and how you want it to work. A few lines is fine.
# CLAUDE.md
## About me
I run [your business]. I use this folder for [what you do].
## How to work
- Be direct and concise
- Don't add things I didn't ask for
- Write findings to a file, not just in the chatStep 3 · Write it down
Keep a project file and close every session
Create a PROJECT.md for each piece of work, and end every session the same way:
Update PROJECT.md with what we did, where things are, what went wrong and what I corrected.Step 4 · Give it access
Connect your first tool
Start the same way as part one: put an export file in the folder. When you're ready to connect a live system, tools plug in through MCP servers. The ones on my tools page install with one command. Google Workspace (calendar, email and Drive) is a common next one. It needs a Google OAuth client first, and the project's README walks you through making one. Then:
claude mcp add google-workspace \
-e GOOGLE_OAUTH_CLIENT_ID=... \
-e GOOGLE_OAUTH_CLIENT_SECRET=... \
-- uvx workspace-mcp --tool-tier coreStep 5 · Write it down
Write a procedure the second time you explain something
When you catch yourself walking it through the same job again, have it write the steps down:
Write the steps we just followed into playbooks/[name].md so you can follow them next time without me explaining.Step 6 · Write it down
Make a rule when the same mistake comes back
The first time it does something you don't like, note it in the project file. The second time, make it a rule: each markdown file in a .claude/rules/ folder loads automatically.
Step 7 · Ask for a blind review
Have a separate agent check the work
Claude Code can start a separate agent that doesn't see your conversation. Before anything ships:
Start a separate agent that can't see this conversation. Give it the finished file and the sources, and have it check every claim. Report problems in order of importance.When it really matters, have another company's model check it too, in ChatGPT or Codex.
Later
When a rule keeps getting broken anyway, stop rewording it. A hook is a small script Claude Code runs before an action, and it can block that action outright. A rule the model can forget isn't a guardrail; a hook is. The guides cover this and more, for people building their own agent setups.
Updated October 2026