I use the claude sonnet model as that is what many people recommend. This model really gives good results for the first 5-15 prompts. After that, it seems to be replaced by something very simple and useless.
LLM is unable to correct even small errors. It just deletes working code. For example, when I tell it that moving a file via drag-n-drop on a form with <input type=“file” /> doesn’t work, it says that we have too many click handlers and composer in response removes half of the click handlers from the whole project. When asked what click handlers have to do with dropping files, the neural network only says sorry.
Even if you still have a project backup, you can’t use neural network queries from this point on, because composer thinks that the deleted lines should be removed.
And it is not always about deletion. Sometimes it happens that the model wants to decorate the interface with various questionable solutions. For example, when you click on a button, shift this button by 20-30 pixels to the right. And you can remove this effect from CSS 5 times, but Composer returns transition to its place at any promt affecting this CSS file.
How do you get out of such situations?