I use AI while learning and building. It can explain an error, help with syntax and give me a first attempt to review. Those are useful things.
I still need to understand what goes into the product. Getting code quickly doesn't tell me whether the feature is correct.
It helps me find a starting point
Sometimes I understand what I want but don't remember the exact syntax. An array method, a regular expression or a configuration example can be a useful prompt.
Then I need to verify the answer. For a library, compare it with the official documentation for the version in the project.
I ask what evidence to look for
Instead of only saying “fix this,” I can ask for possible causes and the checks that would distinguish them.
That gives me somewhere to begin in the console, Network panel, server logs or database. I still have to test each explanation.
Clear repetitive work can be easier to review
Converting an object into a type, drafting test cases or generating sample data can be useful when I can define the result clearly.
The sample data should not contain real private user information. I also need to compare the output with the requirements instead of assuming repetitive work cannot go wrong.
I can ask for another perspective
Where could this fail? What happens if a value is null? Could two requests overlap? Is a permission check missing?
AI can suggest things to investigate. It can also miss a problem or invent one, so a review comment still needs judgment and verification.
I slow down around permissions and secrets
Hiding an admin button doesn't secure an admin endpoint. A database query still needs appropriate access policies. A privileged secret must not be put into public frontend configuration.
With authentication and authorisation, I need to understand who is allowed to do what and where that rule is enforced.
Payments need more than a happy-path example
I want server-side confirmation, clear records and safe handling of retries and duplicates. A pending or unknown result should not be treated as a verified success.
I can't accept a long payment implementation only because it looks organised. The failure paths matter too.
The database needs product decisions first
Can an artist have several profiles? Can a seller own several shops? What happens after deletion? Which fields are public?
AI can help challenge those rules or draft a schema. I need to choose the rules that fit the product before accepting the tables.
I want to be able to maintain what I accept
What enters the function? What does it return or change? How does it fail? Which checks protect the operation?
If I cannot explain those parts, I need to learn more, reduce the example or check the documentation before shipping it.
I want AI to make learning and investigation easier. I still own the decisions, the tests and the maintenance. That responsibility stays with me after the chat is gone.