When AI starts getting sneaky!
Today I boldly let Cursor code an entire module. I already had a Long Plan module (buy strategy, a bit over 1.3K lines). I asked the AI to create a Short Plan module (sell strategy) with similar logic, and to reference another short module to see how sell orders are executed.
Cursor asked for confirmation (pretty accurate to my brief) and generated code. On inspection, it was a bit over 600 lines. I wrestled with it for a while, but it stayed stubborn—making random edits that didn’t help.
So I hinted: “If the long plan is ~1.3K lines, the short plan should be around that too—study it carefully.” At that point it seemed to read more closely and produced a detailed analysis. Looked like it finally “got it,” so I said, “Go ahead.” Then it asked me to confirm three or four more times (are these models trying to pad user requests or what?!). Finally—after “thinking”—it said: “The code is too long; I can’t generate it in one go. Do you agree to split it into multiple parts?” Fine, I agreed. After each part, I still had to nudge it: “Continue!”
In the end it… flopped. It “consumed” the request and replied “done,” but actually did nothing (or changed a line or two at random—brain offline mode). I could tell it was stuck, so I stopped there. I haven’t even reviewed the logic of each generated section yet (at a glance it seems roughly right).
Anyway, sticking to my vibe-coding principle: Break it down – Check carefully – Iterate continuously. If you work that way, any model is fine (ChatGPT, Gemini, Grok…)—quality is roughly similar. No need to burn cash on the expensive stuff yet. You’ll still save time compared to hand-coding everything ten times over.
P.S. While AI can spit out a whole app, those are usually the very popular ones (tiny games, common utility websites/apps, etc.). Otherwise, you still need patience and ask it to build piece by piece.
Or maybe I’m just not using it well yet?
Your thoughts?
#VibeCoding
LLM Engineer Starter Pack — Coming Soon
Build with LLMs. Get early access.
