If you can use a typed language, you want guardrails. Unit tests aren't enough you want integration tests with headless browsers if applicable. It can automatically catch many of its own mistakes earlier.
Be opinionated on your rules files for how you want it to organize the data and code (I built a scaffold template I use all over https://github.com/victusfate/scaffold, fork it make it your own)
If it edits tests you need to push back on why. I'm seeing it too often go to switch case vs simple inheritance. It's too slow sometimes for simple one line change.
When I first adopted AI use I was overly optimistic of it's ability to understand and reliably fix code. One day after several of refactoring, I gave it what I thought was a simple larger task to refactor a single larger source file. I watched in horror as it when completely wrong corrupting two large source files before I terminated it. With undo not an option as many small changes had be made consecutively I ended up having to go back to source control for a clean slate and lost several hours of effort.
I now never trust an LLM to do a task of this size even though I know some people love agents and let them run hog wild on their code bases it will be a cold day in hell before I let even the best frontier model run unsupervised on my code.
There was that time I was sitting in the passenger seat and Microsoft Copilot told us to get off at the wrong exit. Lesson: use Google Maps/Apple Maps/Openstreets instead.
If you can use a typed language, you want guardrails. Unit tests aren't enough you want integration tests with headless browsers if applicable. It can automatically catch many of its own mistakes earlier.
Be opinionated on your rules files for how you want it to organize the data and code (I built a scaffold template I use all over https://github.com/victusfate/scaffold, fork it make it your own)
If it edits tests you need to push back on why. I'm seeing it too often go to switch case vs simple inheritance. It's too slow sometimes for simple one line change.
When I first adopted AI use I was overly optimistic of it's ability to understand and reliably fix code. One day after several of refactoring, I gave it what I thought was a simple larger task to refactor a single larger source file. I watched in horror as it when completely wrong corrupting two large source files before I terminated it. With undo not an option as many small changes had be made consecutively I ended up having to go back to source control for a clean slate and lost several hours of effort.
I now never trust an LLM to do a task of this size even though I know some people love agents and let them run hog wild on their code bases it will be a cold day in hell before I let even the best frontier model run unsupervised on my code.
AI does the first 95% in 3 hours, then leaves you with the last 5% for 3 days.
There was that time I was sitting in the passenger seat and Microsoft Copilot told us to get off at the wrong exit. Lesson: use Google Maps/Apple Maps/Openstreets instead.
[flagged]