Nice troll. Blatantly incorrect. But, I’ll play along in hopes of providing some input to the audience.
AI cannot yet read your mind to guess accurately about anything you are thinking without some kind of input in addition to the code. What about a greenfield project? The code does not yet exist. Upon what, then, would the AI operate without any kind of added prompt instruction?
Of course, though you did not state an qualifiers, you could be thinking of a script that is sufficiently simple to be pasted as the prompt itself. In that case, it may just perform an analysis of it automatically, perhaps a linting, maybe a refactor or bug fix of something that may be obvious. How have you experimented with this idea? Did you even try? What were the results? You could test it easy enough. Do a fresh git clone on a decent sized code base (to avoid any prior AI artifacts interfering with the test.) Then, start claude, cursor, or windsurf/cascade on that fresh code base and just tell the prompt to “proceed” or just “go” and see what happens. It has all of the code. The key detail it does not have is why any of it matters to you. It exists. Yay! All done. lol Please report back after you’ve proven your original conjecture. It may be fascinating to hear of what new ideas the AI has come up with in your behalf.
You seem to be confusing data with an expression of intent. The code base is merely data upon which an AI can operate, inferring some context for the intent expressed by the prompt instructions. Unless there is a glaringly obvious flaw the AI may just happen to find, it cannot guess what needs to be changed. You have to provide AI the intent in addition to the code. You may want behavior X and Y instead of Z. The AI cannot infer that with no other input. Similarly, perhaps you want a better layout of whitespace on the home page. The AI cannot infer that either with no other input. It is very easy for an AI to guess (hallucinate?) incorrectly given poorly expressed or incomplete instructions. Giving even less instructions (none) is not likely to improve the result.
Recently, I used AI (Windsurf, Claude Sonnet 4.5, GPT-5.1) to add an entirely new feature to a product. Prior to this, the feature only existed in my head and in a few conversations I have had with other humans. There is no way an AI could intuit this idea to create it, especially not in the way I envisioned it, at least not without input. Instead, it took several hours of conversations with the AI before it could even document exactly what I wanted created in the way in which I wanted it. I had it create detailed plans for 11 phases of implementation that were suitable for a junior dev. On Friday, I had it implement each phase in turn, so that phases 1 through 9 were completed and working by the end of the day. Phase 10 will be the documentation. Phase 11 will be some polish and final testing. The idea the AI could have come up with and implemented all of this entirely on its own using only the original code as the input is, quite frankly, laughable.
