Cursor, Claude Code or Codex: what should you actually learn?

Which AI coding tool should I learn first? is one of the most common questions people ask us before they start. They have seen a video of someone building an app in an afternoon with Cursor, a thread swearing Claude Code is the only serious option, and an announcement that Codex now comes with their ChatGPT plan. They want to back the right horse before they spend the time.

That answer surprises people, because the internet frames this as a contest with a winner. It is not, for a reason worth understanding before you spend a ringgit: the people asking are rarely the people the tools were built for. Among 113 enrolled Sigmaschool students who filled in our intake survey, 56% had never written a single line of code and 87% rated themselves at level 0 or 1 (n=113, from our confidence gap research). All three tools assume you can read what they write. Most beginners cannot, yet, and no choice of tool fixes that.

So this guide does two jobs. First, a plain description of what each tool actually is, without the benchmark wars. Second, the order we teach them in, taken straight from our own curriculum, and why the order matters more than the pick.

Why the “which is best” debate keeps expiring

Read three comparison articles from six months apart and they will crown three different winners. That is not because the writers are careless. It is because the tools are converging fast. Cursor began as an editor with autocomplete and now runs agents that edit many files and work in the background. Claude Code began in the terminal and now plugs into VS Code and JetBrains and has a desktop app. Codex now spans a command-line tool, an editor extension and cloud tasks that run while you do something else. Whatever one ships, the other two tend to ship something similar within a few months.

Usage reflects that. JetBrains’ January 2026 developer research, as widely reported, put Cursor and Claude Code level at around 18% each of developers using them at work, with GitHub Copilot still ahead. The professionals are split, not converging on a single answer, and many use two tools side by side: an editor for day-to-day changes and a terminal agent for bigger tasks.

The underlying models are a moving target too. Each tool lets you choose between several models, and the leaderboard for which one codes best has changed hands repeatedly. If you pick a tool because of a benchmark score this month, you are choosing on the one thing most certain to change.

Which leaves the useful question: what do you learn by using one of these tools that you keep when you switch to another?

Every item on the right is a fundamentals skill wearing an AI costume. Reviewing a diff is reading code. Describing a task precisely is breaking a problem down, which is what programming always was. Testing on your own machine is operating a development environment. None of it is new, and none of it is taught by the tool. The tool only rewards you for having it.

The clearest example is the rules file. Every serious agent setup now includes a plain text file that tells the AI how your project works: what it is, which commands to run, what it must never touch. Codex and Cursor read a file called AGENTS.md. Claude Code reads CLAUDE.md, which can pull in the AGENTS.md file with a single line. Our own learning platform, Sigmo, does exactly that, so whichever tool a member of our team opens, it reads the same rules.

Where the tool genuinely does matter

It would be convenient to say the choice is irrelevant. It is not, and pretending otherwise would be dishonest.

The interface changes how you learn. An editor like Cursor keeps the code in front of you. Every suggestion appears in the file you are looking at, which makes it easier to notice what changed and why. A terminal agent like Claude Code or Codex works more like a contractor: you brief it, it goes away, it comes back with finished work. That is wonderful when you can judge the work, and quietly harmful when you cannot. For a learner, that difference is bigger than any benchmark.

Teams standardise. If you join a company that runs on Cursor, you are expected to be productive in Cursor from your first week, and an employer may list a tool by name in a job advert. The good news is that the ramp is short if the fundamentals are there. The bad news is that it is very long if they are not, whichever tool you practised on.

Cost adds up for heavy users. Every tool starts around USD 20 a month, but working developers who run agents all day often move to USD 100 to 200 tiers, and the way each one meters usage differs. That is a real decision for a professional. It is not a decision a beginner needs to make.

People sometimes read that sequence as anti-AI. It is the opposite. We hold the agents back for ten missions precisely because we want graduates to be good at directing them, and you cannot direct work you cannot judge. The Mission 10 lessons barely mention which buttons to press. They are about giving the AI the right file, adding a constraint, asking for a plan before any code is written, and reading the plan for the decisions it is unsure about. Those habits work in Copilot, in Cursor, in Claude Code and in Codex, because they are habits of the person, not features of the product.

It is also why students do not pick a favourite tool from a leaderboard. By the time they reach Mission 11, they have opinions from experience, which is the only kind worth having.

If you want to see what happens when this order is reversed, our essay on vibe coding versus fundamentals walks through it. The short version: the project works until the first bug the AI cannot fix, and then the person who built it discovers they did not.

A one-week test, whichever tool you pick

Once you are at stage two or beyond, here is a simple way to find out whether a tool is making you better or just faster. Do it for a week with whatever you chose.

Before each task, write one paragraph describing what done looks like. If you cannot, you are not ready to hand it to an AI, and the tool will happily fill the gap with its own guess.

After each task, explain the diff out loud, line by line, as if to a colleague. Anything you cannot explain, you either learn on the spot or delete. This is the same test our defence interviews run, with the AI switched off.

Once a week, fix a bug with the AI closed. Not because you will work that way, but because it tells you whether the understanding is yours or borrowed. If that week feels impossible, slow down on the agent and spend more time in the editor. Our guide to learning to code with AI without becoming useless has ten more rules in the same spirit.

The fear underneath the question

When we ask enrolled students what they fear most about learning to code, the top answer is not picking the wrong tool, the wrong language or the wrong school. It is their own ability to learn, named by 39% of them (n=62). “Which tool should I learn” is often that fear in a more comfortable shape: if I pick the right one, maybe it will make up for what I do not know yet.

It will not, and that is good news. It means the thing standing between you and directing Cursor, Claude Code or Codex well is not a subscription. It is a few hundred hours of building things yourself, with feedback, until reading code feels ordinary. That is learnable by anyone who puts in the hours, and it keeps paying off whichever tool wins next year. For a wider tour of the landscape, including Copilot and Windsurf, see our roundup of the best AI coding tools.

If you are at stage one, the most useful next step is not a USD 20 subscription. It is finding out whether you enjoy building at all. Our free trial lets you feel the first week of the programme with one signup and no card, and it starts exactly where this guide says to: with your own hands on the keyboard.