Cursor(version 0.42.x) to refactor a TypeScript middleware that handled JWT validation. Everything was fine until I tried to implement a custom caching layer for the user sessions.
I kept feeding the AI the same three files. Every time I asked it to fix a type mismatch in the cache service, it would rewrite the middleware and introduce a bug I had already solved ten minutes prior.
The error was always the same: TypeError: Cannot read properties of undefined (reading 'exp')
.
It was a classic case of context drift. The LLM was hallucinating a version of my User
interface that didn't exist in the current file, but existed in a cached snippet from a previous prompt. I was stuck in a loop where the AI was correcting its own mistakes by creating new ones.
The "Everything is Context" Trap #
Most people treat the @
symbol in Cursor like a magic wand. You @Files
or @Codebase
, and you expect the model to magically know exactly which line of which file is the source of truth.
But here is the reality: the more context you shove into the prompt, the more likely the model is to lose the "signal" in the "noise." I had my entire /services
folder indexed. The model was seeing three different versions of a Session
type across different files and it started guessing.
I tried the standard "forget everything we just talked about" prompt. It didn't work. The model just apologized and then repeated the exact same mistake.
How I actually stopped the bleeding #
I realized I was treating the AI like a human who reads the whole project, rather than a window of tokens. To break the loop, I had to be aggressive about context management.
First, I cleared the chat entirely. Starting a fresh thread is the only way to kill the ghost of previous hallucinations. Then, instead of @Codebase
, I used a very specific set of constraints.
Here is the exact prompt structure that fixed the TypeError
:
Ignore all previous assumptions about the User interface.
Reference ONLY @user.service.ts for the type definition.
The error is specifically on line 42 of @auth.middleware.ts.
Fix the null check for 'exp' without changing the return type of the function.
By forcing the model to ignore its internal "memory" and rely strictly on a single file, the fix happened in one shot.
| Approach | Result | Time to Fix |
| :--- | :--- | :--- |
| @Codebase
(General) | Hallucinated types / Regression | 120 mins |
| @Files
(Multiple) | Confused between versions | 60 mins |
| Strict File Pinning + Clear Chat | Correct implementation | 2 mins |
Why solo prompting is a dead end #
The wild part is that I spent those four hours thinking I was the only one hitting this "context loop." It felt like a bug in Cursor's indexing.
It wasn't. It was a gap in my understanding of how the model weights context.
This is why I started hanging out in an AI Developer Community. When you're coding alone with an LLM, you don't know if the bug is in your code, the model's logic, or the tool's retrieval system. You just end up arguing with a chatbot.
When I posted this specific issue on PromptCube, a few other devs jumped in. One mentioned that they hit the same thing with Claude 3.5 Sonnet when dealing with large monorepos. They pointed out that the "Composer" mode (Cmd+I) handles context differently than the chat sidebar (Cmd+L).
If you're tired of guessing why your AI is suddenly acting stupid, you should join the AI Coding circles. It's the difference between trial-and-error and actually having a strategy.
Hard-won Cursor tips for the sane developer #
If you want to avoid the "debugging the AI" rabbit hole, stop being polite to the LLM and start being restrictive.
1. Kill the "Chat Bloat"
If a thread goes over 10-15 messages and you're starting to see regressions, kill it. Start a new chat. The token window is a sliding scale; the further you get from the original requirement, the more "drift" occurs.
2. Use .cursorrules
for global constraints
I stopped telling the AI "don't use any external libraries for this" in every single prompt. I put it in a
.cursorrules
file at the root of my project. Now it knows my architectural preferences without me wasting tokens on it.### 3. Be a context minimalist
Stop using
@Codebase
for specific logic fixes. It's great for "Where is the login logic located?" but it's terrible for "Fix this specific type error." Use specific file references. Actually, I've found that sharing the exact prompt that triggered the fix is more valuable than the fix itself. That's why Prompt Sharing is such a huge part of the workflow now. You aren't just sharing code; you're sharing the "key" that unlocks the model's correctness.
Rethinking the AI Workflow #
Most developers use AI as a better version of Stack Overflow. That's a mistake. The real power is in the Workflows—the way you pipe context from a bug report to a specific file, then to a targeted prompt, and finally to a verification step.
I used to think that the "best" AI coder was the one who knew the best prompts. I was wrong. The best AI coder is the one who knows exactly how to limit the AI's field of vision so it can't hallucinate.
If you're still fighting with your IDE, stop trying to "prompt engineer" your way out of it in isolation. Join PromptCube. It's a place where we actually tear apart these tools to see where they break and how to make them work for real production code, not just Todo apps.
Next Rise Reforming: Decentralizing Chemical Production →
All Replies （0） #
No replies yet — be the first!