How to use Claude Code: install, setup, first tasks, no coding
How to use Claude Code without coding: which plan you need, step-by-step install, permission modes, four non-code tasks, safety rules and real limits.
A regular Claude chat hands you text. You copy it, paste it somewhere, save the file yourself. Claude Code flips that around. It runs on your computer, sees the folder you point it at, and does the file work itself. You describe the job in plain words, and it reads, edits, creates, renames, and reports back on what it did.
The name confuses people. "Code" here means "runs locally, with your actual files," not "programmers only." Cleaning up a messy folder, pulling data out of a hundred documents, forcing text files into one format: none of that involves writing a line of code.
Which plan you need for Claude Code
The docs are blunt about this. Claude Code requires a Pro, Max, Team, or Enterprise subscription, or a Claude Console account on API billing. The free Claude.ai plan does not include Claude Code access. That is the first thing people trip over. One catch on Team: Claude Code comes with premium seats, not standard ones.
For one person, Pro is the entry point. Anthropic's pricing page lists $20 per month billed monthly, or $200 charged once up front for the year, about $17 per month. Max starts at $100 (5x) and $200 (20x) per month: mostly more headroom on usage limits, plus early access to new models and features. Not needed on day one.
Check the system requirements while you are at it: macOS 13.0+, Windows 10 1809+ or Windows Server 2019+, Ubuntu 20.04+, Debian 10+, or Alpine Linux 3.19+. You also need 4 GB of RAM or more, an x64 or ARM64 processor, internet, and access from a country Anthropic supports. What each plan includes is broken down in our separate piece on Claude pricing.
Installing Claude Code: two routes
Route one, the terminal. One command.
macOS, Linux, WSL:
``
curl -fsSL https://claude.ai/install.sh | bash
``
Windows PowerShell:
``
irm https://claude.ai/install.ps1 | iex
``
The installer pulls in everything it needs; a separate Node.js install only applies to the npm route. On a Mac with Homebrew, brew install --cask claude-code works too. On Windows there is WinGet: winget install Anthropic.ClaudeCode.
Confirm it landed:
``
claude --version
``
You get a version number followed by (Claude Code).
Route two, the app. Claude has a desktop app with a Code tab. Same Claude Code, but with buttons and a diff view of every proposed change, accept or reject on each one. No terminal, no separate CLI install, same plan requirements. If the word "terminal" makes you tense up, start here.
Your first session
Open a terminal in the folder you want to work in and start it:
``
cd /path/to/your/folder
claude
``
On first run it sends you to the browser to log in. Worth knowing in your first five minutes:
Shift+Tabcycles the permission mode/clearstarts a fresh conversation with empty context/compactsqueezes a long conversation into a summary/add-dir pathadds another folder to the session/rewindopens the rewind menu, and so doesEsctwice on an empty prompt
Make your first request a harmless one: "tell me what is in this folder and how it is organized." It reads and answers without changing anything.
Claude Code permission modes: the setting that matters most
Claude Code asks before it edits a file or runs a command. How often depends on the mode. Shift+Tab cycles the three main ones:
- Manual (config value
default; the status-line label varies by CLI version): reads run without asking, along with a built-in list of safe read-only commands such asls,cat, andgit status. Everything else waits for your yes. - Accept edits: file edits inside your working directory apply automatically, and you review afterward. So do file commands in that same directory:
mkdir,touch,rm,rmdir,mv,cp,sed. Those are exactly the ones/rewindcannot undo, so use this mode only on a copy. - Plan: it investigates and proposes a plan without touching your files until you approve it.
A sane starting pattern: begin in Plan, read the plan, approve it, then work in Manual and confirm each step. claude --permission-mode plan starts you there directly.
One mode deserves a warning. bypassPermissions, also reachable through the --dangerously-skip-permissions flag, turns off nearly every confirmation prompt. The docs allow it only in isolated environments like containers or VMs, so keep it off on your everyday laptop.
Four real tasks that have nothing to do with code
Sort out a folder. Copy your Downloads folder somewhere separate, then ask: "group these files by type and year, create subfolders, delete nothing, and give me a list of what moved where." The "delete nothing" and the report request are the load-bearing parts.
Pull data into a table. Point it at a folder of invoices or contracts: "go through every file, extract the date, counterparty, amount, and number, and build a CSV. Where a field is missing, leave it blank and list those files separately." That second sentence matters more: it shows where the tool struggled instead of filling gaps with plausible fiction.
Bulk rename. "Rename every file to 2026-08-02_counterparty_amount, lowercase, spaces replaced with underscores. Show me an old-to-new table first, and only rename after I confirm."
Clean text files by a rule. Forty notes in .txt and .md: "remove the line with the old address from every file and make the headings consistent. Show me the changes for the first three files, then work through the rest quietly and report how many you edited."
The shared trick in all four: show me first, then do it. Claude Code often asks on its own, but writing it in is more reliable. On phrasing so the output matches what you pictured, see fix the prompt first; for more non-technical scenarios, Claude Code without coding.
Safety rules
- Work on a copy. Not your drive root, not your entire Documents folder. One folder for one job.
- Back it up before you start. Not out of fear, but because undo is incomplete, as covered below.
- Move sensitive things out first. Passwords, scanned IDs, and other people's data do not belong in the working folder.
- Read what it is about to run. The permission prompt shows the full command.
- Verify by hand. Count files before and after, open three at random.
Common mistakes and what to check manually
"command not found: claude." It installed, but your terminal has not picked it up yet. Restart the terminal. If the command is still missing, the problem is your PATH or a second install alongside it. Once claude runs, claude doctor reports install health and settings errors.
It is working in the wrong folder. Claude Code sees the folder you launched it in. If it "cannot find" your files, check where you ran cd, or attach the folder with /add-dir.
Undo does not cover everything. /rewind restores edits Claude made with its file editing tools. Files changed by shell commands such as rm, mv, or cp are not tracked, and neither are changes made outside the session. That is what the backup is for.
It described a plan instead of doing the work. You are probably in Plan mode. Look at the mode indicator and press Shift+Tab.
Answers got mushy and it forgot the beginning. A long session eats context. /clear starts over, /compact compresses the history.
The numbers in your table do not add up. Reconcile rows against files. A blank cell is more honest than a neatly filled one.
Where it breaks, and when you need a developer
A scan or a photo-PDF reaches it as an image. The model does see the text, but it slips on poor quality, handwriting, and small print, and a long PDF is read in page ranges, up to 20 pages at a time. So spot-check a batch of scans, and on large volumes run OCR up front.
On big batches it stumbles on outliers. If ten of three hundred files are formatted differently, those ten are your defects, which is why the exceptions list matters more than the summary number. On long jobs it hits your plan's usage limits, and /usage shows where you stand.
Bring in a developer when the cost of a mistake stops being personal: a live customer database, payments, access control, anything in production. Claude Code can write the code. The question is who checks it, and in those places that has to be someone who understands the consequences.
Where to start
Install it, launch it in a test folder, give it one of the four tasks above. Half an hour tells you whether this is your kind of tool.
The AGINE Academy program goes all the way to Claude Code and launching your own products: 81 lessons, 12 blocks, seven tracks, and every lesson ends with something you can actually use. No technical background required. The first four lessons are open with no signup if you want to see how the format works. AGINE Academy is not affiliated with Anthropic.
See also
Questions
Yes. Per the official docs, Claude Code requires a Pro, Max, Team, or Enterprise subscription, or a Claude Console account on API billing. The free Claude.ai plan does not include Claude Code access. For one person the entry point is Pro: $20 per month billed monthly, or $200 charged once up front for the year, about $17 per month. On Team, Claude Code comes with premium seats, not standard ones.
No. You describe tasks in plain language, not in code. Sorting folders, pulling data from documents into a table, bulk renaming, cleaning text files by a rule: none of that requires writing code. The word Code in the name means the tool runs on your own machine with your own files.
Yes, Claude has a desktop app with a Code tab. Same permission modes, plus a diff view with accept and reject buttons. The differences are small: launch flags such as claude --permission-mode plan only exist in the terminal, and project settings are easier to handle there. For your first tasks the app is plenty, and the plan requirement is identical.
Run /rewind, or press Esc twice on an empty prompt, to open the rewind menu. Edited text across forty files with Claude's file editing tools: that comes back. Renamed those same forty files with mv: it does not, because shell commands (rm, mv, cp) are not tracked. Neither are changes made outside the session. Hence the rule: copy the working folder before you start.
Start with the folder: one folder scoped to one job, never your drive root. Then the mode sequence. Plan until you know what it intends to do. Manual when you are walking through steps and confirming each one. Accept edits only on a copy of your data, because it auto-runs mkdir, mv, cp, and rm inside the working folder. Leave bypassPermissions and the --dangerously-skip-permissions flag off: the docs allow those only in isolated containers and VMs.