Claude Code for people who don't code

The name says "code", and that puts a lot of people off. Here's a more useful description: Claude Code is an assistant that can read, write and organize the files in a folder on your computer, and write and run small programs when that's the fastest way to do the job.

If your work lives in files (drafts, spreadsheet exports, interview transcripts, PDFs, a downloads folder you're scared to open), it's for you too. You don't need to read the programs it writes. You do need a safe place for it to work, and a sense of what to hand it.

Where to run it

You have three options, and they share the same engine.

The terminal. Install with one command (on macOS, Linux or WSL: curl -fsSL https://claude.ai/install.sh | bash), then move into your folder and type claude. If you've never opened a terminal, Anthropic has a terminal guide for new users, and there's a walkthrough in your first hour with Claude Code.
The desktop app. The Claude desktop app includes Claude Code as its Code tab, so there's nothing extra to install. Before your first message, pick your project folder with the folder control under the prompt box. It needs a paid plan. The same app also has Cowork, which Anthropic describes as handling longer agentic work; the habits in this post carry over to it.
The web. Claude Code on the web runs in the cloud and works on GitHub repositories. It's handy for developers, less so for a folder of files on your laptop.

Set up a workspace folder

Give Claude one folder to live in, with a shape that makes mistakes cheap:

workspace/
├── CLAUDE.md     notes Claude reads at the start of every session
├── sources/      originals; never edited
├── inbox/        new things to process
├── working/      drafts, scratch files, scripts
└── output/       finished files you'll actually use

CLAUDE.md is a plain text file of house rules that Claude Code loads automatically. Keep it short and specific:

workspace/CLAUDE.md

# My workspace

This folder holds my newsletter research and drafts.

- Never edit files in sources/. Copy them to working/ first.
- Write new files to working/ unless I say otherwise.
- Name files date-first, like 2026-10-07-pricing-notes.md.
- For anything with numbers, write a script, run it, and keep it in working/.
- Ask before deleting, moving or renaming more than five files.
- US spelling, short sentences.

There's a full post on this file: CLAUDE.md: teach your project once.

CLAUDE.md is guidance. Claude follows it well, but nothing forces it. For the one rule that really matters, add a hard rule that Claude Code itself enforces:

workspace/.claude/settings.json

{
  "permissions": {
    "deny": ["Edit(/sources/**)"]
  }
}

Now Claude's editing tools, and the shell commands Claude Code recognizes as file edits, can't touch anything in sources/. A script Claude writes could still change files in ways a rule can't see, which is why the next section exists. The permissions post covers rules in depth.

Safety: copies and version history

Three habits, in order of importance.

Work on copies. Originals go in sources/, and a second copy lives somewhere outside the workspace entirely. If a file matters and exists only once, it doesn't go in the folder yet.

Turn on version history. Git is a free version history tool, and you don't have to learn it to benefit. Your first prompt in a new workspace can be:

Set up git in this folder and commit everything as a starting point. After each task I give you, commit the changes with a short description.

Later you can ask "what changed since this morning?" or "put working/ back the way it was at the last commit". If your workspace syncs to a cloud drive, find out how that service restores old versions before you need it.

Know what rewind can and can't undo. Claude Code checkpoints its own edits. Type /rewind, or press Esc twice with an empty prompt, to roll files back to an earlier point in the session. But files changed only by shell commands, like a script that renames 300 photos, aren't tracked. For bulk operations, git and your backup copy are the real undo button.

While you're learning, stay in a mode that asks before acting. Recent versions often start in auto mode, so in the terminal press Shift+Tab until the status bar says manual mode. Manual mode prompts before edits and commands, and plan mode has Claude explain its plan before it touches anything. See plan mode: think before it types. One more line of caution: check your client contracts before you put their data into any AI tool.

Good first tasks

Start with jobs that are tedious for you, easy to check, and low stakes if they go wrong.

1. Tame a messy folder.

Look through inbox/. Propose a folder structure and date-first file names. Show me the plan as a table and wait for my OK before moving anything.

2. Answer a question from a spreadsheet. Export it to CSV first. Claude Code's file reader refuses Excel files and would have to write a script to open one, while CSV is plain text it reads directly.

Read sources/sales-2026.csv. Which five products grew most from Q2 to Q3? Write the answer to output/q3-growth.md and keep the script you used in working/.

3. Mine your transcripts. Plain text transcripts (.txt, .vtt, .srt) work well.

Read every transcript in sources/interviews/. Pull out each pain point with a direct quote and timestamp, grouped by theme, into working/pain-points.md.

4. Brief a stack of PDFs. Claude reads short PDFs whole and longer ones a few pages at a time.

For each PDF in sources/reports/, write a one-page brief in working/briefs/ with the page number next to every claim.

5. Draft in your voice.

Using working/pain-points.md, outline three newsletter issues. Match the tone of the two examples in sources/past-issues/.

6. Clean up a batch of drafts.

Every file in output/ should start with a title line and a date line. Fix the ones that don't, and list what you changed.

Notice the pattern: say where the input is, what you want, where the output goes, and when to stop and check in.

What it's bad at

Office formatting. It doesn't edit Word or Excel files the way you do. Work in text, Markdown and CSV, and do the final formatting in your usual app.
Being right without checking. It can misread a source or invent a plausible detail. Ask for quotes, page numbers and links, and spot-check them.
Mental math. For anything numerical, ask it to write and run a script and to show you the result. That's why the CLAUDE.md above asks for one.
Huge, vague requests. "Organize my Documents folder" goes worse than "organize inbox/ into five folders and show me the plan first". Point it at specific subfolders.
Your taste, out of the box. Its default voice is generic until you show it examples. Skills are the next step once you know what good looks like.
Things outside the folder. Your Google Drive, Notion or email platform need a connection first. See MCP in plain English.

Your first session

Make the folder. Copy in one real but low-stakes project: last month's interview transcripts, say, or a messy exports folder. Write a five-line CLAUDE.md. Ask Claude to set up git. Then give it task 1 or task 3 from the list above, read what it produces, and tell it what to fix.

That's a full loop: input, instruction, output, review. Once you've done it a few times, save your best prompts as reusable slash commands so the good versions are one word away.