Last week, Cursor faced a very public incident. Users were being logged out when they switched between computers. One of them contacted support and received a courteous reply from "Sam" explaining that this was a new policy: one device per subscription. No such policy existed. Sam was an AI support bot, and it had invented the rule. The incident was widely discussed, some users cancelled their subscriptions, and one of the co-founders issued a public apology. The user received a refund.
We mention this incident because it illustrates the current state of AI tools well. The AI inside the editor is genuinely useful. The AI answering support emails invented a rule and stated it with complete confidence. Both came from the same company, in the same week.
What is Cursor, and why has it been so widely adopted?
Cursor is a code editor built by Anysphere, a company founded in 2022 by a group of MIT students. It is a fork of Visual Studio Code, so anyone who has used VS Code, as most web professionals have, will find it familiar from the first day. Existing extensions and shortcuts continue to work. The addition is a chat panel on the side that can see the entire project.
That panel is what sets it apart. GitHub Copilot mainly completes the line being typed. Cursor answers questions about the codebase, edits several files at once, and now offers an agent mode that makes changes across the project, runs commands and returns a set of edits for the developer to accept or reject.
Developers are also paying for it. Reports this year place Cursor above one hundred million dollars in annual revenue, an unusual figure for a text editor. Developers do not usually pay for editors, which indicates how much time the tool saves them.
What does Cursor change for small web projects?
For large software teams, the main benefit is productivity. For small web projects, in our view, the more significant change is that the editor can now answer questions.
Many small business websites are inherited. One person built a WordPress theme years ago, another added a plugin, and a third modified the footer. No documentation was left behind. Opening such a project used to require an afternoon of reading before any change could be made safely.
With an AI editor, the developer opens the folder and asks questions. Where does the contact form send its emails? Why is this CSS file loaded twice? What does this function do, and what calls it? An answer arrives within seconds. It is not always correct, but it is correct often enough to point to the file that matters.
This alone makes older projects less expensive to fix. For a small business, lower maintenance costs matter more than a faster initial build. Most small websites do not fail because they were poorly built. They fail because each small change costs more than the owner is prepared to pay, so nothing is updated and the site gradually becomes outdated.
Where can Cursor's agent mode cause problems?
Agent mode is also the feature that carries the most risk.
A developer requests one change, and the agent decides along the way that three other elements need improvement. It renames a variable in a file that was not mentioned, reformats a template, or "simplifies" a function that was written awkwardly for a specific reason. Everything still appears correct in the browser. A week later, the French version of a page loses its menu, or an older form stops sending.
There is no ill intent in this behaviour. The agent is an eager assistant with no memory of why things were built the way they were. The same situation occurs with a new human developer, only more slowly.
The second problem is structural. In a project with duplicated folders, old copies of templates and a theme named "theme-final-2", the AI has to guess which version is the active one, and its guesses are only as reliable as the project's organization allows.
The Sam incident also applies here. The support bot story is unrelated to code, but it reflects the same failure: a confident answer, delivered courteously, that no one verified. An editor that invents a function name will at least produce an error. A bot that invents a policy causes a business to lose customers.
Which practices keep AI-assisted changes clean?
Ask for an explanation before any change
On an unfamiliar project, we recommend making the first prompt a question rather than a request: explain how this project is organized, or describe what this file does. If the explanation is incorrect, you learn that the AI misunderstands the project before it has edited anything, at very little cost.
Make small changes and review them
Make one change at a time and read the diff before accepting it. Cursor displays every edit in green and red. When the list is long, it is tempting to select "accept all"; that is precisely when a careful review matters most. If a change is too large to read comfortably, it is too large, and a smaller one should be requested.
Maintain a clean project structure
Delete obsolete copies or, preferably, move them out of the project into an archive folder. Use clear, simple names. Keep a short README that identifies what is live. This is housekeeping, but an AI editor performs considerably better in a well-organized project, as do people.
What should a small business owner know?
Most business owners will never open Cursor, and that is perfectly reasonable. The person who maintains their website may use it, however, and the benefit should be reflected on the invoice. Small fixes that once took an afternoon should now take less time. If they do not, it is reasonable to ask why.
At the same time, faster work does not mean unverified work. We recommend asking the person doing the work whether they review what the AI changed before it goes live. The answer should be yes, every time, without hesitation.
If your website has become one of these inherited projects with little documentation, this is a good year to address it. A proper maintenance clean-up costs considerably less when the tools can read and explain the existing code.