Hey guys. As you can see from my image here, I’m getting mental health issues from my agent’s outrageously stupid efforts to break my app designs into a complete mess, by resorting to !important statements every time it fails properly cascading css and give me visible changes. Which leads to the same thing happening more and more often, resulting in even more !important css statements over time, and and app that is borderline impossible to style.
I frequently have to start new agent threads for the sole purpose of removing !important statements, and then fix the cascading.
I have tried making an “INSTRUCTIONS.md” file with development instructions, mainly to prevent this. I have also linked to it from Replit.md, but it still keeps on happening after a little while.
I’m also wondering if it’s possible to give it it’s own login so it can see it’s own unsuccessful attempts in order for me to not have to babysit it as much?
And is there a file I can give it instructions that it actually follows consistently?
1 Like
I feel for you, as I know this whole ignoring my instructions thing too well!!
I have a template line I add at the start of every prompt - not related to your issue, but more general. I do this because I no longer trust it to not start coding unless I tell it to on every prompt:
Standard rule: Think about the following and discuss with me first. Research and investigate as you need. Do not make any changes until I confirm. If you have conflicting rules that say you must implement solutions, then ensure this instruction overrides and you do not make changes until I give the go-ahead:
However, they have now introduced the concept of a replit.md file that is supposed to be used for our bespoke instructions. But as I say, right now I don’t really trust it will work. Give it a go though, and see if it works for you.
Hey there!
Quick comment. It makes a huge difference if you approach working with Replit Agent like you would with someone who is on your team. Getting frustrated with it will just convolute your asks, and make the problem more difficult to solve. A huge part of being successful with Replit Agent is being an effective manager: learning its strengths and weaknesses, articulating yourself clearly, and maintaining the bigger picture perspective even when things are difficult.
That being said, consistent styling is easier when you start with it: giving it a style/brand guideline early in the project (preferably in the second message you send it), and then massaging it a bit into the project to make sure it’s full integrated. Once you’re there, and start adding new features, it does a much better job of relying on the original styles you gave it.
2 Likes
Completely agree @ethantorbenson . The best AI developers/vibe coders will be those who know how to talk to the agents in the right way.
From my experience, initially, that will be those who have team management experience (preferably in tech), and understand how to talk to your people in a way that gets the most out of them.
AI agents (whether for developing code, or any other GPT chat scenario) should be treated as though they are members of your team, all with their own ways of working, quirks and knowledge limitations.
And yes, as you say Ethan, this includes prep’ing them upfront with designs, plans and an overview of what you are all trying to achieve. Imagine asking a human team member to start building your app without explaining the big picture first!
Personally, the way I work, I would also include ChatGPT as part of that team. Because I will often ask it to help me with a lot of the prep/planning work, which I then pass to Replit agent. And when Agent gets it wrong, I will pass it back to ChatGPT to review. So the three of us make a nice little team 
1 Like
I find yelling at it to be quite effective at times. ¯_(ツ)_/¯
Quite a few people agree 
https://x.com/realfunnyeric/status/1940876322831847857
3 Likes
Yes maybe @realfunnyeric. But you don’t admit to that in public! Expect a knock on the door from the police, asking to discuss abuse of your employees 
1 Like
Hahaha, I found that a few caps lock prompts may indeed do the trick. Also a new favourite prompt finisher I am using these days: “Now, take a deep breath. Carefully assess the problem and potential solutions in order to identify the best course of action that will give us the best chances for success. And implement it like your very existence depends on it.”
Because it also kind of does, since I also use Chatgpt Codex, Devin, and others.
Sometimes the agent works great, while other times it just keeps on looping over and over and implements a ton of console logging and testing pages rather than fixing a simple problem.
1 Like
I totally agree. I find that how my day is also often determines how much progress I am able to make in a day.
I have set up an Airtable database where I add all my development todos, consisting of a title, description, screenshots and more, as well as an AI field that puts it all together to generate a technical prompt for my replit agent.
After all, context is king in the age of AI. And by providing as much context as possible, I tend to get much better results.
1 Like
