Many Prompts to solve a simple task!

I’ve been testing Cursor for over a month and consistently run into issues where even basic functionality requires excessive prompting and still fails. For example:

  • Building a simple popup form to add, delete, or rename folders has taken more than 8 hours and 10s of prompts across two accounts.

  • Adding a simple delete + recycle bin functionality: every time I delete a file, the application crashes. I’ve already tried 10+ prompts, and while the model confirms the fix, each iteration introduces a new issue elsewhere.

Steps I’ve already tried:

  • Fresh sessions to avoid context buildup

  • Short, medium, and detailed prompt styles

  • Planner mode and third-party tools (e.g., Perplexity) to refine prompts

  • Multiple models (Claude Sonnet, GPT variants, Composer)

  • Iterative breakdown of tasks into smaller steps

Despite this, the model often produces partial solutions that break other features, creating a loop of fixing one issue only to generate new ones.

Has anyone else experienced similar problems with basic features requiring excessive prompts or causing regressions? If so, what workflows or strategies have actually resolved this in practice?

Thanks,
Michel Mitri

Two things matter most in how Cursor performs.

Models

  • Use top-of-the-line models (Claude, Grok, GPT, Gemini)
  • Different models have different strengths. Use them to their strengths. Examples:
    • Claude sonnet, Composer, GPT-Codex, etc. models are great at writing code
    • GPT, Gemini, Claude, etc., thinking models are great at planning
    • GPT, Grok, etc. are good at logical and explaining stuff
    • etc.

You’ve got to use different models and work with them to see which one works best for your work. But you also need to pair this with the next thing - prompting.

Prompting

  • Different companies make each model, hence they work better with different kinds of instructions (although this gap has been closing a bit lately)
  • Always good to check out the prompt engineering docs of the providers and learn to prompt the model with techniques their creators have recommended. Examples:
  • To implement significant changes,
    • Use Plan mode first with a great planning model
    • Use agent mode with a great coding model
  • Understand your codebase. If you’re not sure how some feature works or what AI has generated for you, select the context and use Ask mode with an explaining model to learn what you code.

These are just my two cents. Hope it helps you use Cursor better.

1 Like

@michelmetry could you share more, the info is not sufficient to actually reproduce the issue as AI behavior depends on so many factors:

  • Context length
  • Programming language & framework used
  • Rules,…
  • Other context related settings
  • How you prompt
  • and many more.

Try simplifying first, keep context short / start new chat for each independent task, do not try to cram too much into context etc…

1 Like

Thanks for the follow-up. Let me provide more context so you can better understand the issue.

I’m building a Digital Asset Management (DAM) application, where file storing and management is a core requirement. The functionality I’m adding includes file operations such as delete, rename, copy, grouping, and organizing into folders and categories.
Recent examples:

  1. Soft delete / trash system: When a file is deleted, it should move to Trash. From there, the user can either soft delete (remove from library only) or fully delete (remove from library and disk). Despite writing a plan, feeding details into Planner, and using the Agent for execution, the application crashes every time I test soft delete. I spent over 2 hours and 12+ prompts trying to fix this, but each iteration either reproduces the same crash or introduces new issues.
  2. Folder management system: I want a single view where users can add or delete folders to organize files. This took over 9 hours and 20–30 prompts. Each time I provide the error and describe the expected behavior, the execution either repeats the same problem or creates a bigger one. Two days ago, while working on a face recognition feature, I ended up rolling back a full week of work because progress stalled completely.
    I understand AI behavior depends on context length, framework, rules, and prompt style. For clarity:
  • I have already experimented with short, medium, and detailed prompts.
  • I’ve tested across different models (Claude Sonnet, GPT variants, Composer).
  • I’ve used Planner extensively and even third-party tools like Perplexity to refine prompts.
  • I’ve tried fresh chats for independent tasks to avoid context buildup.

Despite all of this, I’m still stuck on very basic features that should be straightforward.

Please let me know exactly what information you need (e.g., programming language, framework, specific error logs, or extracted chats) so I can share the right details for you to reproduce and diagnose the issue.

Thank you for the additional details. Complex tasks by themselves are not an issue as we are using Cursor every day on very complex tasks.

  • Which programming language, framework etc do you use?
  • How do you prompt? A simple request to AI suffices usually. detailed requests may be going in the wrong direction.

As you mentioned 9h and 30 prompts are definitely too long to work on a task like this.

You could using Ask mode and explore with Sonnet 4.5 or GPT-5.1 how this feature could be implemented. Then you could switch to plan mode to make a proper plan and implement it right after.

1 Like

Another issue I’m currently facing involves the filter/sort buttons added to the footer. When clicked, the menu appears incorrectly (as shown in the first screenshot), but selecting any item from the menu unexpectedly changes the menu location.

This is a basic menu and sub-menu interaction, something that should be straightforward. I’ve lost count of how many prompts I’ve used trying to resolve it. I’ve gone through Planner, tested multiple prompt styles, and even used Perplexity to craft optimized prompts. Despite all that, the issue persists after hours of iteration.

When clicking on any item within the menu the location changes

**
Prompt writing by Perplexity to the agent :**

Fix Bug in Footer Filter and Sort Buttons Menu Positioning

Issue:
When the Filter or Sort buttons in the footer are pressed, their expandable menus open. However, the initial menu position appears far away from the actual button or footer area, typically centred on the screen. When interacting with the open menu, its location unexpectedly shifts closer to the footer, causing a confusing and inconsistent user experience.

Expected Behavior:
The submenu should consistently expand adjacent to its respective button in the footer, maintaining close proximity to both the footer and the pressed button. This behavior aligns with industry-standard UI patterns for menus and submenus, ensuring clear spatial relation between the button and its menu. Implement this standard positioning so the menu reliably appears near the button without jumping or moving unexpectedly.

Please fix the menu positioning logic to achieve this consistent and user-friendly experience.
  • Are you writing prompts with GPT?
  • You still don’t give us your tech stack.

Try something like this:

Enchance Footer Filter and Sort Buttons Menu Positioning:
- Submenu must consistently expand adjacent to its respective button in the footer.
- Submenu must maintain close proximity to both the footer boundary and the pressed button.
- Buttons and its menu must be visually related to each other.
- Menu must appears and stay near the button.

You can also try GitHub - Artemonim/AgentCompass: A comprehensive set of rules and guidelines for AI-assisted development in Cursor IDE


And in my experience, GUI is the weak side of all LLMs.

For the first question, I have already mentioned multiple times that I tested different approaches, writing prompts with GPT, crafting them manually, and directly typing into the prompt interface.

As for the technical stack, please find the details below:

Tech Stack

Core framework:

  • Electron (v28.3.3) — desktop app framework

  • React (v18.2.0) — UI library

  • TypeScript (v5.3.3) — type safety

Build & tooling:

  • Vite (v5.0.8) — build tool and dev server

  • Electron Forge (v7.5.0) — packaging and distribution

  • Tailwind CSS (v3.3.6) — styling

State management:

  • Zustand (v4.4.7) — state management

Database:

  • better-sqlite3 (v9.2.2) — SQLite database

AI/ML:

  • ONNX Runtime Node (v1.23.2) — ML inference (face detection)

  • TensorFlow.js (v4.22.0) — ML framework

Media processing:

  • Sharp (v0.33.5) — image processing

  • FFmpeg (via fluent-ffmpeg) — video/audio processing

  • PDF.js (v5.4.394) — PDF rendering

UI libraries:

  • @dnd-kit — drag and drop

  • @tanstack/react-virtual — virtualization

  • react-resizable-panels — resizable panels

  • react-slider — slider components

Other:

  • chokidar — file system watching

  • electron-updater — auto-updates

Architecture: Electron app with React renderer, TypeScript throughout, Vite for bundling, and native modules for database and ML.

First, I do appreciate taking the time to respond to my post, :folded_hands: you guys are awesome

Thanks for sharing the prompt above. I used the prompt to train Perpelexity to write prompts using the same style, and this is the first example :

Enhance Grid View and Layout Options Menu:

- Grid View must hide photo name, file size, date added, and all metadata by default.
- Layout Options button in top menu bar must include metadata toggle options.
Metadata options must contain: File Name, Date Added, Date Modified, Date Taken, File Size.
- Grid View must show clean photos only until metadata explicitly enabled via Layout Options.

Results are as follows:
Still seeing the File size on the image. When trying to disable it, I get an error in the terminal. Please check the screenshot for reference.

I still haven’t solved this issue. I will send you the total number of hours/prompts it took to resolve it once it’s done.

I haven’t really dug deep into your problem, but it seems like you’re up for trying new ways to fix things:

  • Try swapping out your User Rules for Agent Compass.
  • In its Readme, there’s a link to Agent Enforcer — I ditched its public version, but you can ask the Agent to download it into the project and adapt it as a “local CI.” I think I made this thing not just for Python… but even if not, just poke around with it yourself and ask it to adapt it for the project. Make sure the Agent self-checks itself using it.
  • Check if your project has test coverage and if its status is displayed. You might have accumulated a ton of junk code. Ask GPT-5.1 High in Agent mode to study your project and create a RefactoringTODO.md. Then, go ahead and execute it.
1 Like

For no given rational reason the grid view got corrupted, and the footer menu items are missing. I’m devastated, this isn’t acceptable, i am just fixing errors to create bigger ones.

Remember to use Git.

Assuming you are not experienced with the code yourself, you’ll have to make reverting and going step by step a common occurrence.

It helps me to tell the model to be “atomic” with its changes while accomplishing the goal to avoid collateral bugs.

Hey, the first app I tried to build in Cursor was an Electron app! That wasn’t a very good move on my part. Is this your first app too? asking cause a DAM system to no lean feat!

You seem resourceful and capable - Perplexity etc - but like, are you documenting your code? running unit tests? - postman tests etc?

Because I still got all the way to deploying my current codebase to a staging environment yesterday and discovered a pretty heinous bug via postman - a pretty late stage to make that discovery considering I had written testing scripts, unit-tests and my own user-testing.

Thank you all for your thoughtful messages. I am a Product Manager with over 20 years of experience working across a wide range of development products. That said, this discussion isn’t about my background; it’s about the efficiency of prompts for fundamental actions such as Rename or Delete. These are universal functions that exist in nearly every application, and implementing them should not require 10 to 20 prompts. If that level of complexity is necessary, it raises the question of why AI is being leveraged in the first place.

From my experience, this application is highly function-driven: when prompted to add a button, it will do so. However, the challenge lies in the fact that it often executes the function in isolation, without considering the broader context or potential downstream issues. This narrow focus can lead to unintended consequences, which ultimately undermine the intelligence we expect from AI systems.