I am just a single retired guy trying to create a SEO generator for a website that I have and I was trying to do it with ChatGPT and that’s when I found out about hallucination then I learned about Cursor and when I first got on it, it fixed everything that I’ve been working on for almost 2 months in about a day and a half, and then after that, it just starts screwing everything up, changing things that I didn’t even want changed and keeps adding to it but never fixing the problem. I tried logging out logging back in using multiple agents different AI’s. I’ve even tried using a different user to see if that would make a difference. I’m just going in circles and sliding down a hill and I’m at the point of just giving up because one step forward nine steps backwards I need, help any suggestions? I’ve tried almost everything.
Do you code at all? What is the project state like? Do you manually review the code AI generates? Does not seem like an issue with Cursor by your description.
So, “green field”, a programmer term for “starting from nothing”, is relatively easy. When you have nothing in place, there is nothing to screw up, so initial rounds of implementing some new project with an AI Agent, usually go very well.
This, sadly, can be a bit deceptive to non-technical types, and give some…perhaps, unrealistic expectations. Once you HAVE SOMETHING, then the subsequent work to refine, fix, enhance, and otherwise expand upon the original work, becomes significantly more complex. Its complex, by nature…it does not really matter if you are doing the work yourself (i.e. as a programmer manually writing all the code) or using an agent. There is just complexity in the task of REFINING EXISTING CODE…
An AI Agent CAN do a good job, however, usually they will not do a good job without guardrails. What are guardrails? Rules. Plans. Details. Explicit instructions. SCOPE-LIMITED TARGETED CHANGES.
As a non-technical person, it IS possible to manage your own software project. However, you will need to learn some things, notably about how to manage, and control, an AI Agent. Without guardrails, without some degree of proper control…well, LLMs are highly non-deterministic tools. So yes, they can, and WILL, run roughshod over your now-existing code, and WREAK HAVOC. It does not matter, really, what model you use. GPT, Claude, Grok, Gemini, etc. They all have slightly different capabilities, strengths and weaknesses…but, at this stage of LLM design and capabilities overall, they are ALL subject to the same general weaknesses, and the main one is their probabilistic nature, which leads to non-deterministic results (without effective guardrails that is.)
The problem is not really the agent, or cursor, or even the code. These are aspects of the problem, to a degree. The main problem though, is the lack of guardrails. So my recommendation is, do some research, learn how Cursor Rules, Agent Skills, Custom Commands, Plan mode, Debug mode, and a few other features of Cursor are designed and intended to work. Do some research in each of these areas, to see how other people are using these features.
Good thing is, you can instruct the Cursor agent, to create rules, custom commands, even agent skills, FOR YOU. But, you should understand how they work, and how they help to manage the Agent/Model, and keep it doing things the way you want it to, the way you need it to, to keep it from “going off the rails” and ruining your code, etc.
I think this is an essential skill for ANYONE using an agent to build software, but if you are non-technical, I think that learning how to create and maintain the essential guardrails around the agent/model, is one of the single most important things. Especially if your project/product is not a “one-shot wonder”…where it can be built in one shot from nothing to working, which is usually only the case for the simplest of projects. Learn how to use the tool, and the tool will do its job better. It won’t resolve everything overnight…but, you should, over time, go from +1-9, to +1-7, +2-5, +3-4, then eventually you’ll really start moving!
Hope that helps.
- Replace your User Rules with GitHub - Artemonim/AgentCompass: A comprehensive set of rules and guidelines for AI-assisted development in Cursor IDE
- Ask GPT-5.2 XHigh or Sonnet/Opus 4.5 to create local CI system based on GitHub - Artemonim/AgentEnforcer: Powerful and flexible code quality checking tool designed to be used by both developers and AI agents. (without MCP support and only for the project tech stack)
- Ask same model to create
RefactoringTODO.mdin the root of the project - Proceed + Force Agent to use local CI in the end of working
- Give me 2 Github stars if it helps
Replace your User Rules with GitHub - Artemonim/AgentCompass: A comprehensive set of rules and guidelines for AI-assisted development in Cursor IDE
Thanks for the link. Interested in seeing how you have set up your rules. I’ve developed a fairly large set of rules rather organically over time here, and they have definitely helped corral the agent/models and keep them on track, but I suspect its probably time to go through, and refine, and better organize, what I’ve put together.
To my surprise, I haven’t touched the core ruleset I wrote two months into Cursor for a long time. I only add additional rules to work around Cursor bugs and project-specific rules within the repositories themselves.
Larger doesn’t mean better. I tried to make the rules as unambiguous and compact as possible.
Large set, meaning, many rules, in many files. For a lot of different things, spanning…a dozen repos. 
Hey! I see that @jrista and @Artemonim have already shared great tips. Here are links to the official docs that should help:
Key concepts for controlling the Agent:
-
Cursor Rules: create rules for your project: Rules | Cursor Docs
- Rules help the agent understand your project context and avoid unwanted changes
- Start with simple rules about what it should NOT change
-
Plan Mode (Shift+Tab): use it for complex tasks: Planning | Cursor Docs
- The agent will first create a plan you can edit before it runs anything
- This gives you control over what will be done
-
Ask Mode: use it to learn the code without making changes: Modes | Cursor Docs
- The agent only reads and answers, it won’t change anything
Practical tips:
- Make small, specific requests instead of one big task
- Use Plan Mode to review the approach before running it
- Create a simple
.cursorrulesfile with basic instructions for your project
If you want help with a specific issue, share what the agent is changing wrong and what you want it to do instead.
No, I do not code at all I can recognize some code just from doing this, but overall no I don’t code. I had an idea for an SEO generator and it started off really well until I started tweaking little things here and there and then the AI start hallucinating changing things that wasn’t requested and I just went totally backwards from there and haven’t caught up to where I was.
I appreciate your advice and I’ll definitely look into that, your explanation sounded exactly what happened. It started off perfect until I start tweaking it and then flew right off the edge no guard rails.
Yes then this is the perfect example of the vibecoding pitfalls. When you start creating, you start with a simple idea. Then you want to extend it and that is where the bad code LLMs produced start being an issue.
It is difficult to manage if you do not code and do not review everything the model does. LLMs struggle with structuring code and handling abstractions. LLM will do the bare minimum by default and it also produces code that is nearly impossible to extend.
So bottom line I need to spend a couple years learning code to actually get a working project.
It is possible to learn how to guide an LLM throughout the project creation. Set guidelines. Force proper code structure by instructions. Review every change using a different LLM. Spend a lot of time focused on ensuring the project structure rather than making new features.
Thank you vary much for this, it does really helps.
by the way i had issues when i started, started readsing a bunch and i now do pretty good. you do not need years to learn now days. you need patience and willing to ask people questions. thats what i did.
you need to learn how the llm reads you project and prompt. then you will understand how to reveiw and also (mainly how to manage your Ai.
Not necessarily. You need to understand concepts, and need to understand how the agent and model operate. The key is that a model, the LLM (Large Language Model) is a probabilistic type of computational resource. That means that unlike most computer programs, which follow very strict instructions and follow very explicit code paths, to get from INPUT to OUTPUT, and produce the same results each time you execute with the same input (what we call DETERMINISTIC; there are some caveats, such as when involving timestamps, random input, etc.), an LLM can follow many different paths to get from INPUT to OUTPUT, and will not always produce the same result. Run the same query (prompt) again and again and again, without any modifications, with a clean/fresh/new context, and an LLM will produce output that will differ to one degree or another, and may differ quite wildly. As such, LLMs produce NON-DETERMINISTIC results.
To some degree, the non-deterministic nature of LLMs can be leveraged for positive benefit. Using an agent to develop a program, you can get some fairly “creative” results, as well as “insightful” results, as a consequence of the models probabilistic nature. I try to leverage this with web app/mobile app design and implementation a lot. There are some really interesting solutions that the agent can come up with, when it comes to UI design. I have my own design ideas, but there are times when the agent comes up with something I hadn’t thought of, that is more interesting or more dynamic or with better animations or something, than my own ideas.
When it comes to the critical functional/behavioral code, though, the non-deterministic nature of an agent, as you have experienced, can be quite detrimental. You do not necessarily need to know how to code…and to be quite frank, two years is probably not enough, if becoming a programmer is not really your goal, as this is a very large, very complex and very challenging field. What you need to learn, is how to MANAGE THE AGENT. There are concepts you need to learn about how LLMs work, about how an Agent works, about how Cursor works, and about how software projects are designed. You can learn higher level abstract concepts about software, without necessarily needing to know a lot about code. Your product will be….lets just say “messy”, and over time the mess will present its own challenges. However with proper management of the agent, and an understanding of how software projects work, you should be able to keep the agent on track well enough to produce a product, bring it to market, and see if you can build a business around it. If you CAN, and can bring in some revenue, then at that point you could think about brining on a proper technical resource to manage the agent, clean up the messy code, and maintain the project longer term.
Actually, over the past few days, auto performance is really off. It is not doing what it used to do. Its like not reading the code its working with, or its not going off of its index. Something broke or they degraded the performance of it or something.
its probably the latter, a money issue, it might be using worse LLMs so they can recoup money