note 03
Pick an agent
A coding agent can read your files, write code, run commands, and keep going until the thing works — or until it gets stuck. You talk to it in ordinary language. You stay in charge. Pick one and finish something before you collect more.
01
The three you will hear about
They all do the same job: sit in a project and change it with you. They are different shapes. You do not need all three.
This site is written around Cursor because that is what we use, not because the others are worse. The habit is the same everywhere: say what you want, watch what changed, click it, send it back.
Cursor
An editor with the agent built in. You see the files, you click around, the agent edits in place. One window for the code and the conversation. The easiest first pick if you want to see the project while you talk.
Claude Code
Anthropic’s coding agent. You run it in a folder in the terminal. Strong at long, careful work across a whole project. Same company as Claude Desktop — different app. Do not mix them up.
Codex
OpenAI’s coding agent. Also lives in the terminal (and it shows up inside ChatGPT). Same job as Claude Code: sit in the project and change files. Different model family, different taste.
02
Others that exist
You will also hear GitHub Copilot, Windsurf, Gemini CLI, Aider, Cline, Continue, Amp, OpenCode, and a new one every few months. Some live in an editor. Some live in the terminal. Some live in the cloud and you watch a replay.
They are real tools. You do not need to try them. Pick Cursor, Claude Code, or Codex. Finish a site. Then get curious if you want.
03
Chat apps are not a coding setup
Claude Desktop, the ChatGPT app, and the Gemini app are good for talking: “what should I build,” “explain this error,” “write me a paragraph.” They answer in a bubble. If they give you code, you copy it into the project yourself. The chat and the files on disk drift apart. That gets messy fast.
Claude Desktop is not Claude Code. ChatGPT on your phone is not Codex in your project folder. Same companies, different jobs.
Cursor is a desktop app too — but it is an editor. The files on disk are the source of truth. That is the difference that matters.
04
Why the terminal is better for building
The project is a folder of files. A terminal agent works in that folder. You see every command it runs. You can stop it with Ctrl+C. You can run the same commands later yourself. Git, the package installer, the dev server — those are terminal tools. Agents are better when they can use them in front of you, not behind a chat bubble.
A chat app hides the computer. The computer is the point.
If you picked Claude Code or Codex, the terminal is where they live. If you picked Cursor, it has a terminal panel inside the editor — same commands. When you want the agent or the server in a window you cannot lose behind a tab, use a real terminal. There is a whole note for that, including why iTerm2 and a cheat sheet.
05
Models inside Cursor
Cursor can call lots of models. The picker is usually near the chat. Auto means Cursor chooses for you. That is the right setting for most work — changing a page, fixing a button, adding a section. Stay on Auto until you have a reason to leave.
Leave Auto when the same task keeps failing, when you want a second opinion from a different family, when the change is big and you want it to think longer, or when you are curious. Do not switch models every message. If one family keeps failing, try a different family — not a slightly different size of the same one.
Families, not brand names: Claude (Anthropic) is usually careful and good at following a messy project. GPT and Codex (OpenAI) are often fast and direct. Gemini (Google) is another family — useful when the first two disagree. Cursor also has its own models, built for the editor: often cheaper and quick. Names on the list will change. The families stay.
Some models have an effort or thinking control — how hard they think before they start editing. Higher effort is slower, more thorough, and costs more. Use it when you are stuck or the change is scary. Do not turn it up to rewrite a heading.
Once you have an agent, leave it a note. AGENTS.md is a file in your project that tells the helper who is in charge and how you like to work. There is a whole note with a starter you can copy and change.
Pick one agent. Finish something. Switching tools feels like progress and usually isn’t. Auto is boring on purpose.