Claude Code for non-developers: getting started

Claude Code is a terminal-based AI coding assistant from Anthropic that reads your repo, answers questions about it, and edits files when you ask. Non-developers (founders, PMs, ops folks) use it to navigate codebases, draft PRDs grounded in real code, and ask "where is X defined" without bothering an engineer. You install it with one npm command, point it at a folder, and start typing in plain English.

If you have ever opened a GitHub repo, scrolled for 30 seconds, and closed the tab, this guide is for you. We will get Claude Code installed, walk through the first five prompts that pay for themselves, set up a personal-context file so it answers like it knows your business, and lay out the safety rails so you do not accidentally push a broken commit at 11pm.

What Claude Code actually is (and isn't)

Claude Code is a CLI (command-line interface) tool. You open a terminal, type claude, and a chat session starts inside the folder you ran it from. It can read every file in that folder, write new files, edit existing files, and run shell commands when you give it permission.

It is not Cursor. Cursor is an editor (a fork of VS Code). Claude Code lives in the terminal and has no GUI. That sounds intimidating, but for non-developers it is actually easier: there is one input box and one output stream. You do not need to know what a sidebar, a command palette, or a gutter is.

It is also not ChatGPT. ChatGPT does not have your repo. Claude Code does. That means it can answer "where does the trial conversion logic live?" with a real file path and line number, not a guess.

Why non-developers should care in 2026

Three things changed between 2024 and 2026 that make Claude Code usable by people who do not write code daily.

First, the models got better at reading large repos without losing the plot. Claude Sonnet 4.5 handles 1M-token contexts cleanly, so a 200-file Next.js codebase fits in one prompt.

Second, the CLI added a permission model. Every shell command and file write asks you to approve it. You can read the entire repo without ever giving it write access if that makes you more comfortable.

Third, the CLAUDE.md convention turned context into a file you can edit. Drop your business rules, pricing tiers, and "never touch this folder" notes into one markdown file, and every session inherits them. We use this for everything from explaining our AI-native engineering hiring rubric to flagging which routes are billing-critical.

Install in 90 seconds

You need three things: a terminal, Node.js 18+, and an Anthropic API key (or a Claude Pro/Max subscription).

On a Mac, open the Terminal app (Spotlight, type "terminal"). On Windows, install Windows Terminal from the Microsoft Store and use PowerShell.

Run this:

npm install -g @anthropic-ai/claude-code

Then cd into a project folder. If you do not have one, clone any public repo:

git clone https://github.com/vercel/next.js
cd next.js
claude

The first run will ask you to log in. If you have Claude Max ($200/month, our recommendation for heavy use), pick that. If you do not, paste an API key from console.anthropic.com.

You will see a prompt like > waiting for input. That's it. You are in.

The first five prompts that earn back the install

Most non-developers freeze at the empty prompt. Here are the five we hand to every founder on their first session.

1. "Give me a 10-line summary of what this codebase does." This forces Claude Code to scan the file tree, read the README, and synthesize. You will know in 30 seconds if your engineer's "it's a standard Next.js app" was accurate or hiding three side projects.

2. "Where is the user authentication logic? Give me file paths." This is the "where is X defined" prompt. Replace "user authentication" with whatever business concept you care about: billing, onboarding, the email sender, the retry queue. You get back something like lib/auth/session.ts:42 and can open that file (or just ask Claude to explain it).

3. "Draft a PRD for adding [feature]. Read the existing code and tell me what would need to change." This is the killer prompt for PMs. Claude Code reads the relevant files, then writes a PRD that names actual functions and routes that need touching. No more PRDs that an engineer rewrites because they assumed the wrong architecture.

4. "What does this folder do?" Run it with cd into any subfolder. Useful when an engineer says "the issue is in the worker package" and you want to understand the worker package without a 30-minute walkthrough.

5. "Find every place we hardcode a price. List them with line numbers." Replace "price" with any constant you care about. Great for ops folks running pricing changes, copy audits, or compliance sweeps.

CLAUDE.md: your personal-context file

The single highest-impact thing a non-developer can do is write a CLAUDE.md file at the root of the repo. Claude Code reads it automatically at the start of every session.

A useful CLAUDE.md for a founder looks like this:

# Project context for Claude Code

## What we are
Cadence is an on-demand engineering marketplace. Founders book vetted engineers by the week.

## Pricing (never invent tiers)
- Junior $500/week
- Mid $1,000/week
- Senior $1,500/week
- Lead $2,000/week

## Folders to never modify
- /infra (Terraform, ops owns this)
- /data/snapshots (production exports)

## Who I am
I am the founder, not an engineer. When you suggest code changes,
explain the trade-offs in plain English first. Do not run migrations
or push commits without asking.

Three things make this work. The "what we are" section keeps Claude from suggesting features that fight your business model. The "folders to never modify" section is a soft guardrail (Claude respects it). The "who I am" section changes the explanation style and adds a personal safety rail.

You can layer CLAUDE.md files. One at the repo root for company-wide rules, one per package for technical rules, one in your home directory (~/.claude/CLAUDE.md) for personal preferences that apply to every project.

Five workflows that actually pay back

Researching a codebase before a vendor meeting. Before a call with a Stripe integration partner, spend 10 minutes asking Claude Code to map every place your app touches Stripe. Walk into the call knowing your own code better than the vendor.

Drafting a PRD grounded in real code. Ask Claude to write the PRD, then ask it to list the open questions an engineer would have. You get a PRD that is 80% done before the engineer sees it. Pair this with our AI-assisted refactoring playbook if the PRD involves rewriting an existing feature.

Answering "is this hard?" When an engineer says a request will take two weeks, ask Claude Code "if I wanted to do X, what files would change and roughly how much code?" You will not get a perfect estimate, but you will know whether two weeks is a 1x or 10x overestimate.

Auditing for compliance or copy changes. "Find every email template that mentions the old company name." "List every API endpoint that does not require authentication." These take an engineer half a day. They take Claude 90 seconds.

Translating engineer-speak. Paste a Slack message from your CTO and ask "explain this to me in plain English, then suggest two questions I should ask back." This is the most-used prompt at our portfolio companies.

Safety rails (read this before you panic)

Claude Code can edit files and run shell commands. By default it asks before doing either, but if you click through prompts on autopilot, you can break things. Five rails keep you safe.

1. Run it on a branch, not main. Before starting, run git checkout -b claude-experiments. If anything goes wrong, git checkout main puts you back.

2. Never approve commands that contain rm -rf, DROP TABLE, or --force. If Claude proposes one, ask "why?" before approving. Usually there is a safer alternative.

3. Use read-only mode for exploration. Start sessions with claude --read-only when you are just learning the codebase. It can answer questions but cannot edit anything.

4. Don't approve commits. Let Claude propose changes; you (or an engineer) approve and commit. The git commit step is your circuit breaker.

5. Keep a CLAUDE.md "do not touch" list. Mark database migration files, payment code, and infra configs as off-limits. Claude will still read them to answer questions, but it will flag any edit as a violation.

If you want to go deeper on the agentic side, our guide on Cursor's agent mode in production covers the same patterns for editor-based workflows.

Claude Code vs Cursor vs ChatGPT for non-developers

These three tools get confused constantly. Here is the honest comparison.

Tool	Best for non-devs	Sees your repo	Edits files	Learning curve	Monthly cost
Claude Code	Repo Q&A, PRDs from real code, audits	Yes (the folder you run it in)	Yes, with per-action approval	Low (one input box)	$20 Pro / $200 Max
Cursor	Pair-programming with an engineer, AI-assisted edits	Yes (open folder)	Yes, with diff review	Medium (full VS Code surface)	$20 Pro
ChatGPT	Brainstorming, snippet help, no-context Q&A	No (unless you paste)	No	Lowest (web chat)	$20 Plus

Where Cursor wins: visual diffs, multi-tab editing, debugging with breakpoints, the full IDE experience if you ever want to write code yourself.

Where ChatGPT wins: zero setup, works on a phone, fine for a quick "what does this error mean" question.

Where Claude Code wins for non-devs: it is the only one of the three that is comfortable reading a 200-file repo end to end and answering business questions about it. For founders and PMs, that is the entire job.

If you are deciding which tool to invest in first, our build-vs-buy decider walks through the choice with a few questions about your workflow.

What to do this week

Block 30 minutes. Install Claude Code. Point it at your company's main repo. Write a 20-line CLAUDE.md. Run the five starter prompts above. Total cost: half an hour and $20 if you do not already have Claude Pro.

If those 30 minutes deliver value (they will), the next move is to pair this with an AI-native engineer who can actually ship the changes Claude surfaces. Every engineer on Cadence is AI-native by default, vetted on Cursor, Claude Code, and Copilot fluency before they unlock bookings. Our pool has roughly 12,800 vetted engineers, and the median time to first commit on a fresh booking is 27 hours. You can book a mid engineer ($1,000/week) for a 48-hour trial and see the difference between an engineer who runs Claude Code daily and one who treats it as autocomplete.

For the deeper economic argument on why this matters now, the 2026 AI engineering ROI numbers lay out the dollar math for disciplined teams.

If you want a 60-second recommendation on whether your next feature should be built, bought, or booked out, our Build/Buy/Book decider takes three questions and gives you a routing call.

FAQ

Do I need to know how to code to use Claude Code?

No. You need to know how to open a terminal and type plain English. The five starter prompts above are entirely in English. The most-used prompts at portfolio companies are "explain this to me" and "where is X defined", neither of which require coding knowledge.

Is Claude Code safe to run on my production codebase?

Yes, if you follow two rules: run it on a branch (not main), and never approve commits or shell commands you do not understand. Default mode asks before every file write and every shell command. For maximum safety, use read-only mode (claude --read-only) when exploring.

How is Claude Code different from Cursor?

Cursor is a code editor (a fork of VS Code) with AI baked in. Claude Code is a terminal CLI. Cursor is better if you actually edit code; Claude Code is better if you want to ask questions about a codebase without opening an editor. Many engineers use both: Cursor for editing, Claude Code for repo-wide research.

What does CLAUDE.md do?

It is a markdown file at the root of your repo that Claude Code reads automatically at the start of every session. Put your business context, off-limits folders, pricing rules, and personal preferences in it. Every session inherits the context without you having to repeat yourself.

Can a non-technical founder actually evaluate AI-native engineering skills?

Yes, with the right rubric. Ask candidates to share their CLAUDE.md file and walk you through a recent multi-step prompt ladder they ran. If they cannot show you either, they are using AI as autocomplete, not as a working partner. Cadence's voice interview scores exactly this, and our breakdown of AI-assisted technical interviews covers the full rubric if you want to run it yourself.