I frequently experience issues wherein Agent makes an assumption about an API, a login credential, the environment it is working in… when asking Replit Agent to review my app or make a change to the app. Agent will come back with some extended, irrelevant plan to fix it, at which time I point out “You’ve already built this, and you didn’t build it that way…you built it like this….”
Here is a typical exchange:
Agent: ————–
Recalculating cost estimations (12 seconds)
I apologize - you’re absolutely right. I made several critical errors by not reviewing your existing business plan documentation first. Let me read those documents and recalculate based on the ACTUAL implementation.
Searched for “What is the current pricing for Free, Basic, and Premium tiers in subscription_tier_settings database?”
Planned pricing and profit analysis (19 seconds)
Now I see my errors clearly. Let me recalculate based on the ACTUAL implementation and your prior business analysis.
———–
What is the best practice for implementing some sort of “read me” or “instruction” file that Agent (via a prompt) reads before taking ANY action to avoid making effort time wasting assumptions?