Cursor, Claude Code and Codex: how they work and how we use them
These are the tools behind the websites we build. They are not website builders. They are ways of working on a real codebase with an AI that can read every file, make changes across all of them, run the tests and read the errors. Here is what each one does and where it sits in a build.
The idea
An agent does not answer. It acts.
The jump from a chat window to these tools is not that the model got smarter. It is that it was given hands: permission to look at your actual project and change it, then check its own work.
- 01
It reads the real codebase
Not a snippet you pasted. It searches the project, opens the files that matter and sees how things are already done, so a change fits the code around it instead of inventing a second way to do the same thing.
- 02
It makes the change across every file
Renaming a concept might touch a dozen files. An agent does all of them in one pass, which is the part that used to be slow, dull and error-prone.
- 03
It runs things and reads the output
The type checker, the test suite, the build, the linter. This is the step that matters most: the agent sees its own failures and fixes them before a human ever looks.
- 04
A person reviews and decides
The diff gets read, questioned and often rejected. This is the step vibe coding skips, and it is the entire difference between a tool and a liability.
The three
What each one is for.
They overlap, most teams use more than one, and the differences are about where you sit while you work rather than what they can do.
Cursor
A code editor, built on the same foundation as VS Code, with AI woven through it. You are in the file, watching changes appear, accepting or rejecting them as you go. Best when you know roughly what you want and want to stay close to the code: fast edits, refactors, and completions that understand the whole project.
Claude Code
Anthropic's agentic coding tool, which works in the terminal, inside an editor or on the web. You describe a task rather than an edit, and it plans, works across the project, runs commands and reports back. Best for the jobs that span many files or need a lot of reading first: a migration, a refactor, tracking down why something fails.
Codex
OpenAI's coding agent, available as a command-line tool and as a cloud agent that works on a task in the background. Useful for handing off well-defined, self-contained work and collecting the result, and as a second opinion when another tool has been going in circles.
Why more than one
They are not interchangeable in practice. Models have different habits, and a problem one gets stuck on is often obvious to another. Using two is a cheap way to catch a confident mistake, which is the failure mode that matters. Knowing which one to reach for, and when all of them are wrong, is the part that took five years.
In a build
Where they sit in a seven-day website.
The tools compress the building. They do not compress the deciding, and the deciding is what makes a website work.
- 01
Monday is a conversation, not a tool
Who the site is for, what expensive problem the business solves, what an enquiry is worth, and what this business can claim that its competitors cannot. No AI involved, because a model can only return the average of what exists and the whole point of this half day is to find what does not.
- 02
Structure and copy come next
What pages exist, what each one has to do, what it says. AI drafts and argues with us; the decisions are ours and the client's, and the words are shaped by people who spoke to them. Five years of watching which pages earned enquiries and which got skipped is what settles the arguments.
- 03
The build is where the agents earn their keep
Components, pages, forms, responsive behaviour and the wiring between them. Days of typing become hours, and the time saved goes into review and the details nobody prompts for.
- 04
The checks are not optional
Every build is measured before it goes live: performance on a mobile connection, accessibility to WCAG 2.2 AA, metadata and structured data, real form delivery, and behaviour analytics installed and confirmed. An agent helps fix what these find. It does not get to decide they passed.
FAQ
Questions people ask about this
What is Cursor and what is it used for?
Cursor is an AI-powered code editor built on the same open-source foundation as VS Code. It adds completions that understand your whole project, an inline edit mode for changing code by describing the change, and an agent mode that can work across multiple files. Developers use it as their day-to-day editor rather than as a website builder: you stay in the code, and the AI works alongside you rather than instead of you.
What is Claude Code?
Claude Code is Anthropic's agentic coding tool. Rather than suggesting completions in an editor, it takes a described task, explores the codebase, makes changes across as many files as needed, runs commands such as tests and builds, reads the results and iterates. It runs in a terminal, inside an editor, or in a browser. It suits work that requires understanding a project before changing it.
What is OpenAI Codex?
Codex is OpenAI's coding agent, available as a command-line tool and as a cloud agent that works on a task independently and reports back. The name has history: OpenAI previously used it for the model behind early GitHub Copilot. Today it refers to the agent product, used for handing off self-contained tasks and reviewing the result.
Which AI coding tool is best?
For staying in the code and editing as you go, an AI editor like Cursor. For tasks that span a whole project and need exploration first, an agent like Claude Code. For handing off well-defined work to run in the background, a cloud agent like Codex. Most professional teams use more than one, because a problem one model gets stuck on is frequently obvious to another, and disagreement between them is a useful signal.
Are these tools the same as Lovable or Bolt?
No. Lovable and Bolt are hosted platforms that generate and run an application for you, aimed at people who want a result without touching code. Cursor, Claude Code and Codex work on a codebase that lives in your own repository, on your own hosting, in a framework any developer can pick up. They assume someone will read the code, which is precisely why we use them for client work.
Does using AI make a website cheaper?
It makes the building cheaper, which is why a complete website can be a fixed price with a short timeline rather than an open-ended project. It does not make the strategy, the copy, the review or the accountability cheaper, because those are people, and they are the parts that decide whether the site stands out or reads like every other one. Our prices reflect that split: what AI compressed got cheaper, and what is judgement did not.
Start your brief
Start your brief
Tell us what you're launching and what it needs to do. We reply within one business day with a fixed price and the first available Monday.
- Websites from $15,000 ex GST, fixed. The exact figure comes with the reply.
- A reply within one business day, from Michael, not a form robot.
- A thirty-minute call to hear what the site has to do.
- A fixed-price scope, sent as a web page with a version number.
- Kick-off on the next available Monday, held for you. We take four studio builds a month.