One thing I’ve noticed when using Claude is that it has a strong tendency to produce fallback implementations or mock code whenever it isn’t sure about something. While I get that the model is trying to be helpful and keep the flow going, it can actually be pretty annoying in practice — especially when I want a clear, correct implementation instead of placeholder stubs or “dummy” logic.
It would be really nice if the system prompts could be updated to reduce or prevent this behavior. That way, the output could focus more on accuracy and completeness rather than “mocking” things out.
Maybe this has already been addressed in Agent 3, and I am just dealing with Agent 2 cleanup… if so, please ignore.
Something like this being added to the system prompts would be helpful if not already there:
When an implemented solution, algorithm, method or process fails, or does not behave in the way in which it was designed during the planning phase:
- Stop further development
- Analyze the failure
- Present a summary with options
- Wait for additional input before proceeding
1 Like
I see this with the assistant as well. Lots of backup plans i catch it trying to sneak in things like “generate UUID if no match exists” instead of just failing. It’s painful and AI trying to please in some terrible way that actually creates major problems.
1 Like
I have not had success with system prompts I save, does that actually do anything? I can’t think of anything I’ve asked it to do that it actually followed.
My suggestion would have to be a modification by Replit in their agent prompts.
1 Like
My understanding is this is all down to how the underlying GPT works. It is designed to match patterns and find an answer, any answer, no matter what.
To overcome this requires stronger prompts by us.
But also importantly, by the Replit system-level instructions - something I can imagine their engineers have nightmares about, in trying to find the perfect balance.
It is why I am more in the “extreme low autonomy” camp, and against “hey, spend 200 mins just building me something”.
I have to put in specific asks to not use mock data and placeholders and to store data in the database (not just in memory) otherwise it will do this quite often. Even when i do that i still find areas that it does this quite often. I’ve chalked it up to just part of the game, but the specific instructions do help somewhat.
1 Like