Once we let AI tools run on their own, either as agents or in more manual patterns, we eventually run into the scenario where the AI has removed parts of the code that it deemed unnecessary for the current task & context, or may have started to completely rewrite working parts that sometimes weren't related to it's input.
So much, that it's bubbled up to be a meme, and people are coming up with workarounds.
But this raises a valid question when working with AI tools:
Whose code has the authority?
...
Outdated context history
The AI tool overwriting intentional user code, is not only a problem when letting an AI run on code it hasn't touched before, but can also happen when interacting with tools like Cursor's composer, where the knowledge the AI has of the files it works on & the ground truth of these files can be different, when the user has made manual changes to the code, the AI doesn't yet have in it's context.
So when Composer runs it's "Apply" function for it's own generated code, it assumes this code it hasn't seen before, is outdated & should be removed. Especially on small value changes.
This is fixable on the integration side, through reloading the current files on each request, but for now it requires a system rule to prompt the AI to respect new code as intentional.
Do not touch code unrelated to your current task without user confirmation.
If you're encountering new code when applying changes, assume they are intentional & do not remove them.
What about comments?
While many models are already good at distinguishing between documenting commments & commented out code, not all commented out code is to be removed. Yet some definitely is.
In conditional cases like this, a simple system rule could prevent the AI from completing wanted clean up tasks. For scenarios like these, I propose the instruction scratchpad feature.
WIP
In most cases the users code should be given authority & assume it is intentional.
Removing code proactively has a high chance of missing intentions.