An idea gets a working first version
People use “vibecoding” for slightly different things. Here, think of building by describing what you want to an AI, trying the result and steering the next change. You spend less time typing each line and more time saying “nearly, but that button should actually do this”.
For a small experiment, that is genuinely exciting. A utility you kept putting off, an interface idea, an audio experiment: you can get something tangible quickly enough to discover whether it is worth pursuing. The first version is a question you can test, not a finished product you have to defend.
Working once is a very small test
The trap is mistaking a convincing screen for a reliable system. Try the wrong input, an expired session, a failed network request and an account that should not see another user’s data. Ask where secrets live and what is logged. If you cannot explain those parts, keep real customer data out of it.
Work in small changes you can review and undo. Keep version control, run tests and read the relevant diff, even if the same agent says everything passed. Let it help write the tests, but make sure the expected behaviour comes from the requirement, not from whatever it happened to implement.
Give the tool a sensible workspace
An agent does not need production administrator access to move a heading. Use a disposable or isolated development environment, limited credentials and explicit approval for consequential actions. Check the actual permissions of the tool you chose; a friendly chat interface can still run very real commands.
I like the speed. I also want to understand the thing well enough to maintain it when the happy path ends. The useful handover is not “the AI built it”. It is “here is what it does, how we checked it, and how we roll it back”.