l2silver

Using chatgpt (paid) and Claude (free):

When I try to code a somewhat complex project (backend with login capabilities, jwt, refreshtoken, and e2e tests) and I put all the requirements in the initial prompt, the output is usually condensed and incomplete, like it has a maximum output length that it can't exceed.

When I give it small prompts, I get more verbose and complete answers, but the prompts seem to forget the previous prompts, and throw out logic that was written previously.

What's the best way to feed instructions to these AI systems for relatively complex projects?

Were there software developer skills that he used to evade capture, or is it more or less trivial to pull off this kind of attack for regular people?