Guide · 8 of 8

Going further

You can build a thing, save it, put it online and give it data. This last chapter is the extra kit: features of Claude Code and Codex you'll grow into, how the limits feel day to day, how to stay safe, and where to ask when you're stuck. Read it now for the shape, and come back when you need a piece.

Claude Code

Subagents

A subagent is a helper that Claude Code starts for a side task. It works in its own space, then reports back a short answer, so your main session doesn't fill up with its reading. You don't need to set anything up. Just ask:

Use a subagent to read every file in the levels folder and tell me which levels have no exit. Don't change anything.

Good for looking things up across many files, or checking work while the main session keeps going.

Skills and slash commands

Slash commands are the things you type starting with /, like /help, /clear and /model. Type / to see the list. A skill is a saved set of instructions for a job you do often, like "how we add a new level" or "how to write a project page". Claude Code loads it when the task matches, or when you call it by name. You can ask Claude Code to write one for you from a job you just did together.

Hooks

A hook is a command that runs by itself at a set moment, for example after every file edit, or before the agent runs a command. You'd use one to tidy the code after each edit, or to stop certain commands from ever running. Hooks live in Claude Code's settings. The agent can set one up if you say plainly what you want: "after every edit, run the formatter".

MCP: plugging in other tools

MCP (Model Context Protocol) is a standard way to connect the agent to other tools: a browser it can drive, a database it can query, your GitHub issues, a design file. Each tool comes as an MCP server you add once. After that the agent can use it like any other tool. Only add servers from people you trust, because the agent will act through them.

Codex

Cloud tasks. Codex can run a task on OpenAI's computers instead of yours, with codex cloud or at chatgpt.com/codex. You hand it a job, close the laptop, and look at the result later.
/review. Inside Codex, /review reads your changes and points out problems. Run it before you commit something big. It is a second pair of eyes, not a guarantee.
The IDE extension. Codex also runs inside your code editor, so you can chat with it next to your files. There's a desktop app too.

Using both on the same repo

You can use Claude Code and Codex on one project. Claude Code reads CLAUDE.md; Codex reads AGENTS.md. Put the same text in both, so they follow the same rules. This guide's repo does exactly that.

Run one at a time on the same files, and commit between them. Then if one makes a mess, git checkout -- . takes you back (see chapter 4). A handy pattern: build with one, then ask the other to review.

Costs and limits in practice

Both tools come with your plan, and both have limits: a rolling 5-hour window plus a weekly cap. In practice that means you can work steadily for a good while, then it tells you you've hit the limit. You wait for the reset, or switch to the other tool. Big asks use it up faster than small ones.

In Codex, /status shows where you are. Plans and prices change, so check claude.com/pricing and chatgpt.com/pricing rather than trusting any number you read elsewhere.

Habits that make your allowance go further:

Smaller asks. One feature at a time. You were doing this anyway (rule 4).
Clear between tasks. In Claude Code, type /clear when you start something unrelated, so it isn't rereading the old conversation. /compact shrinks a long conversation when you want to keep going. In Codex, start a fresh session.
Don't paste huge files. The agent can read files itself. Say "look at game.js" instead of pasting all of it.

Staying safe

The agent can run commands on your computer. Most of the time that's the point. A few habits keep it from going wrong.

Don't run a command you don't understand. If the agent asks to run something, or a web page tells you to paste something into the terminal, ask the agent first: "what does this command do, and is it safe?"
Never give an agent your passwords. Not your email, not your bank, not your Cloudflare or GitHub password. Logins happen in the browser, where you type them.
Be careful with rm. It deletes files, and there is no bin to fish them out of. If you see rm, especially with -rf, read the path twice before you say yes.
Back up with git. Commit when things work, and push to GitHub. Then the worst case is losing an hour, not a project.

Get a code editor, when you want one

You have built things without one. You will know you want an editor when you keep wanting to see two files at once, read a change line by line before saying yes, or fix a word yourself instead of asking. We use VS Code (Visual Studio Code). It is free and works on all three systems.

Go to code.visualstudio.com and download it. On Windows, install the Windows version, even if you use WSL. It talks to Ubuntu for you.
Install it like any other app. On a Mac, drag it into Applications.
Mac only: open VS Code, press Cmd + Shift + P, type shell command, and pick "Shell Command: Install 'code' command in PATH". This lets you open VS Code from the terminal.
In the terminal, go to a project folder and open it. The dot means "this folder".
```
code .
```
You should see: VS Code opens with your files down the left. It may ask if you trust the authors of the files in this folder. You are the author. Click Yes.

On WSL, the first code . takes a minute while VS Code sets itself up to work with Ubuntu. Let it finish. Both agents also run inside VS Code as extensions, so you can chat next to your files.

Where to ask for help

Ask the agent. "I'm stuck on this, here's the error, what are my options?" works more often than you'd think.
Claude Code docs: code.claude.com/docs.
Free course, Claude Code 101: academy.claude.com.
Codex docs: learn.chatgpt.com/docs/codex/cli.
GitHub's command-line tool: cli.github.com.
Cloudflare docs: developers.cloudflare.com.

What you just did

You met subagents, skills, hooks and MCP in Claude Code, and cloud tasks, /review and the IDE extension in Codex. You know how to share one repo between both, how the limits feel, and the safety habits that matter.

Now build something

That's the whole guide. The projects start with a one-file game and go up to multiplayer and real databases. Pick one that looks fun, paste the brief, and go.

Make things that are yours. A game your friends play at lunch. A tracker that fits how you actually live. A page with your name on it. You described it, you checked it, you fixed it, and it's online. That counts.