All lessons · 44
Your first coding agent: the smallest useful setup
Pick a coding agent — Claude Code, Codex CLI or opencode, two of them free to start — install it, sign in, and give it a first safe task: a small web page in a fresh folder. No programming experience needed, about 15 minutes, and you finish with a working result.
A coding agent is an AI that doesn’t just answer — it acts: creates files, writes code, checks its own work, and reports back. That is the “implementer” role the second lesson warns about: the work is fine to hand over, the understanding is not. So keep the oracle stance from that lesson — the agent implements, you stay the one who decides and understands. Here is the smallest setup that actually does something useful — a small website, in a folder, on your own laptop.
Pick your agent: three good options
One tool, one command, one login — and you pick which one. All three below do the same job: they read your folder, write files, run commands and ask before anything risky. What differs is what they cost to start:
| Agent | Starts free? | Command you’ll type | Pick it if… |
|---|---|---|---|
| Claude Code · install | No — needs a paid plan | claude |
You already pay for Claude, or want the most polished agent and don’t mind a subscription |
| Codex CLI | Yes — free ChatGPT plan, with usage limits | codex |
You want to spend nothing today and already have a ChatGPT account |
| opencode · install | Yes — open source, free models included, or bring any model’s key | opencode |
You want no lock-in: your pick of model, including free and local ones |
They all work on Windows, macOS and Linux, and all three ask before they do anything risky — which is exactly why a non-programmer can drive them safely.
You have already used a chat assistant: you ask, it answers in text. What does a coding agent add on top of that?
pick an answer — you get the explanation right away
The honest costs: time, and possibly money
- Time — about 15 minutes, most of it waiting for installs.
- Money — and here the three tools are not equal. Claude Code is excellent — in my experience the most polished coding agent you can install today — but it is not free. It needs a paid Claude plan (or pay-as-you-go API credit) before it will do a single task; the free plan does not include it. The entry-level plan is enough for an evening of playing. Plans and prices change often, so check the current pricing before you sign up, and prefer a subscription over a card on a meter: a fixed monthly fee has a ceiling, a per-token meter does not.
If you are not ready to pay, don’t close this page. Take Codex CLI or opencode instead: both install and run for free within their limits (Codex on a free ChatGPT plan, opencode with the free models it ships with), and everything below works the same for all three. You can move to Claude Code later — the habit you build here is the same habit.
Not a terminal person at all? Take the desktop app
Every one of the three also ships as a normal window with buttons. Same tool, no typed commands — download, install, sign in, pick a folder:
- Claude Desktop — claude.ai/download. Paid plan required, as with the terminal version.
- ChatGPT desktop app (Codex) — developers.openai.com/codex/app. Works on a free ChatGPT plan, with limits.
- opencode desktop — opencode.ai/download. Free models included; connect one and choose a folder.
We go through the terminal version below because it makes one thing obvious: the folder you start the agent in is its working area, and anything outside it takes your permission. That rule — one project, one folder, plus reading what it asks for — is the safety story whichever interface you pick, and it applies to every agent on the market.
The setup, step by step
Open your terminal
Everything below happens in the terminal window from the previous lesson. Windows: Start → “Terminal”. macOS: Cmd+Space → “Terminal”. Linux: Ctrl+Alt+T.
Install your agent
Run the one command that matches your tool and your system — press the copy button, paste it into the terminal, press Enter, and wait. Text will scroll by; that’s the installer talking.
Claude Code — official install guide. macOS, Linux or WSL:
Claude Code · macOS, Linux, WSLcurl -fsSL https://claude.ai/install.sh | bashClaude Code, Windows PowerShell (the prompt starts with
PS):Claude Code · Windows PowerShellirm https://claude.ai/install.ps1 | iexCodex CLI — official install guide. macOS or Linux:
Codex CLI · macOS, Linuxcurl -fsSL https://chatgpt.com/codex/install.sh | shCodex CLI, Windows PowerShell:
Codex CLI · Windows PowerShellpowershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"Codex CLI, any system with Node.js:
Codex CLI · npmnpm install -g @openai/codexCodex CLI, macOS with Homebrew:
Codex CLI · Homebrewbrew install --cask codexopencode — official install guide. macOS, Linux or WSL:
opencode · macOS, Linux, WSLcurl -fsSL https://opencode.ai/install | bashopencode on Windows, with Chocolatey or Scoop — the docs recommend WSL as the smoother route:
opencode · Windows Chocolateychoco install opencodeopencode · Windows Scoopscoop install opencodeopencode, on any system with Node.js:
opencode · npmnpm install -g opencode-aiWhen it stops and the blinking cursor comes back, it’s done. Errors? Don’t guess — copy the exact text into ChatGPT or Claude and ask what to do. That works more often than you’d think.
Check it installed
Still in the terminal, run the version command of the tool you installed:
claude --versioncodex --versionopencode --versionIf it prints a version number, the tool is on your machine and working. If it says “command not found”, close the terminal, open a fresh one and try again — new programs are often only found by new windows.
Log in
In the same terminal, type your tool’s command (
claude,codexoropencode) and press Enter. The first launch walks you through a login:- Claude Code — choose “sign in with your Claude account” (the same account as on claude.ai), not the API option. This is the step that needs the paid plan.
- Codex — choose “Sign in with ChatGPT”. A free ChatGPT account works; a paid plan just raises the limits.
- opencode — run
/connectinside the app, pick opencode Zen (the team’s own curated models, free ones included), sign in at opencode.ai/auth and paste the key. Or connect OpenRouter or another provider with your own key.
A browser window opens, you log in, done. The app says welcome.
On Windows, one optional extra
Installing Git for Windows is optional, but it gives the agent a proper shell to work in and makes it noticeably more capable. If you skip it, everything here still works.
Make a new empty folder and enter it
This is the “safe zone” rule: by default the agent works inside this folder, and it has to ask before going outside it. Still in the terminal:
mkdir my-first-page cd my-first-pageThen start the agent with its own command —
claude,codexoropencode. A chat interface appears right in the terminal. You’re running an AI that can touch files. Feels big. It is.Give it the first safe task
Paste the prompt from the box below, press Enter, and watch it work: it will create files, write the code, then show you a summary and ask permission for anything that matters.
Look at what it made
In the same folder it created an
index.html. Open that folder in your file explorer and double-click the file — your page opens in the browser. You just shipped software.
If a step doesn't look like the screenshots in some tutorial
These tools update constantly, and so do their install commands. If a command here is rejected, open the tool’s official install page and copy the command from there — the shape of what follows (install, verify, log in, work inside one folder) stays the same.
The loop you just started
Every round of an agent's work is a fresh request to the model. Halfway through a session, what is in that request?
pick an answer — you get the explanation right away
Every coding agent runs the same cycle. Knowing it makes the whole thing predictable:
One plain-language request. Not code — the result you want.
The better you describe the destination, the fewer round trips it takes.
The first prompt (copy it)
Make me a small personal page in this folder as index.html.
About me: [name, what you do, one interesting fact]
The page needs: a short intro, 3 facts about me, and my contact email [email].
Style: clean and modern, dark background, one accent color of your choice.
Then explain to me in 3 sentences what files you created and what each does.That last paragraph — “explain what you created” — is the oracle mindset again. The agent does the work, you keep the understanding.
What a first session actually looks like
1$ claude --version2x.y.z (Claude Code)3 4$ claude5 6> Make me a small personal page as index.html.7 8● Creating index.html...9● Writing HTML + CSS... done.10 11Done. Created:12- index.html (your page: header, 3 facts, contacts)13 14Open it in your browser to see the result.15 16> make the accent color red instead17● Updating index.html... done. Accent color changed to red.18 19> /exit20Bye.See the rhythm: you ask, it acts, you review, you refine. Ten rounds like this and you have a real page — and a real skill. Whichever of the three you installed, the transcript looks the same; only the program name and the colors change.
Rules of the safe road
-
One folder per project. Always start the agent inside a fresh, empty folder. Never run it in “Documents” or at the top of your disk. Deleting the whole experiment is then one keypress.
-
Read before you approve. The agent asks before significant actions. The permission dialog is not a formality — it’s the steering wheel. Here is one, so you know what it looks like before you meet it for real:
The agent asking firstpress the button below to send the first messagesent to the modelnothing yet — every message is sent in fullPress send next → and watch the middle of the transcript: the agent stops and asks before it touches anything, and nothing moves until you answer — it asks before acting outside its folder, and before anything destructive inside it, like this delete. Notice the two outcomes: say no, and the files stay exactly as they were; say yes, and the change lands. Now read that middle button carefully: “yes, and don’t ask again” is not the same as “yes” — it is yes plus a standing permission, so the next action of that kind happens without asking you. Say yes often. Say “don’t ask again” only when you would be fine never being asked about that exact action again.
-
Wrong result? Just say so. “The email section is missing, add it” is a complete instruction. No restarting, no reinstalling.
-
Lost? Exit and re-enter. Type
/exit(or press Ctrl+D) to quit — in some tools it’s/quitor pressing Ctrl+C twice. To stop something mid-run, press Ctrl+C: that cancels the current action, it doesn’t quit the agent. Worst case, close the terminal window; nothing is lost. -
Hit a limit? That’s the plan, not a bug. On a free plan you will eventually get “come back later” or a smaller model at busy times. Switch to a lighter model, or wait it out.
-
It can be wrong about its own work. The page might not open, a feature might be missing. That’s normal and fixable: tell it the exact symptom, or paste the error back.
-
Want to keep going? The natural next step is publishing that page for free — it’s its own article: How to create and host a website for free.
The takeaway
One tool installed, one login, one folder, one prompt — and you’re no longer someone who “doesn’t do computers”. Everything you build from here is the same loop: describe, approve, review, refine.
What makes a coding agent different from a chat assistant?
Where should you start the agent for your first project?
The install just finished. Which command proves it actually works?
You don't want to pay anything yet. Which of these can you actually start with today?
The agent shows a permission dialog before a significant action. What is the right reaction?
Questions?
Set it up? Broke something? Both are great stories — bring them to the Telegram channel.