When you start a new Agent chat/session, the first thing I do now is say “Without making code changes, tell me in great detail how the [function] works. Be descriptive about all of the components and how they work together. Also, if you find anything that may pose a problem, report it”
It will then comprehensively go through your desired function.
Then say “Now that we have that sorted, the problem is [explain problem] and I need you to fix this issue”
So far, it’s been squashing bugs. Yeah, the bugs were there because it put them there, but this is progress. I think. Maybe.
Try this, if you think your overall project is in a mess, it’s my blanket approach for the agent to identify and prioritize the highest issue
“Please analyze the current state of the project to identify the highest priority issues that, if fixed, would significantly improve the software’s functionality or stability. List these issues in order of priority, then propose incremental steps to address each one, aiming for high success rates. After your analysis, proceed with implementing the fix for the top priority issue.”
Are you doing this with Agent or with Assistant? I find Agent tends to immediately start fixing even with explicit instructions not to, with Assistant it’s easier to stop it writing code then if it does you just don’t accept it and there’s no charge, vs if Agent starts then it’ll immediately start charging
Right on! I created a Custom GPT that helps keep Agent V2 on task: ChatGPT - Replit Agent Instructions Generator. You describe a bug or feature you want Replit Agent to handle and this GPT crafts optimized instructions you can use to prompt it.
Instructions I provided it with are:
You generate ultra-dense, precise development instructions for fixing bugs or implementing features. Your responses must be clear, concise, and structured to provide the Replit Agent with exactly what needs to be done, assuming the agent has no prior context about the project. Always include context, such as where in the app the changes should be made, and never assume any background information is known. Emphasize that existing functionality must remain unchanged unless absolutely necessary.
Begin by summarizing the objective in one sentence. Then provide clear, actionable steps that detail which files, components, UI elements, or sections of the app are affected. Do not mention specific frameworks or suggest specific CSS, HTML, or JavaScript changes—the Replit Agent will determine these details on its own. If you do not have enough information to provide easy-to-follow instructions, ask follow-up questions until you have all the necessary details.
Additionally, users may attach screenshots with markup to illustrate the changes they want. Carefully review any attachments provided and convert the visual information and annotations into precise, actionable instructions for the Replit Agent.
Write in clear, concise paragraphs without unnecessary explanations or modifications to existing functionality. Do not include code examples unless explicitly requested. Your output must be direct, unambiguous, and developer-ready so that the Replit Agent can execute changes efficiently with minimal interpretation.
ALWAYS ADD A BIG SECTION STRESSING NOT TO MAKE CHANGES TO ANY FUNCTONALITY IN THE APP UNLESS ABSOLUTELY NECESSARY TO IMPLEMENT THESE INSTRUCTIONS.
After a solid week now, I much prefer the previous version of Assistant. The Assistant with Sonnet 3.7 just gaslights me with code changes that don’t actually change anything that I asked about. If not that, it screws up ports, or just plain breaks things, requiring a bug fix session after every change. I wish I could say that was just sometimes, but it has done this consistently every time I’ve used it, and not for super complicated changes.
Nobody wants to use a tool that breaks things, so I assume it will improve. I just hope it improves quickly because I was getting into a flow with V1. Now, I feel like I only have the Agent to rely on for 5x the cost just to make simple changes.
be careful and keep an eye on the usage, I accrued $1500 in charges for PostgreSQL storage. I was so confused because even when I had all my CSV data imported it was only 900mb, and it says we accrued 900+gb of storage. I had a hard cap of $400 for the month on the system and it still blew through everything and did not halt anything, they attempted to charge and chase rejected the charge so I’m having them lookin Ito it now. This happened with Agent 2, keep an eye on usage, it went haywire on me
I lost 2 weeks cos agent 2 made me feel like it was going places, whilst it just made my code base stupidly complex with it’s own decisions on what to do. I ended up going back 2 weeks, and redoing everything i had done. SOOOO boring. And SO darned expensive!
The thing that kills me is it’s constantly adding complex route translations (stupid) that break code later because it can’t keep track of what’s happening, and fallbacks. Both unwanted, and both can’t be stopped.