Comparing AI Coding Platforms: Cursor, V0, Bolt.new, and Windsurf for Key Visual Generation System

With the rapid advancement of foundational models like VLMs (e.g., Minimax, Qwen-V) and cutting-edge image-generation technologies such as Flux 1.1, we’re entering a new era of creative possibilities. Coupled with models like T5 for enhanced understanding of text prompts in latent space, these tools enable the production of advertising-grade Key Visuals (KVs) with remarkable realism. At our organization, we’ve tested the potential of combining these models to generate high-quality KVs. My task was to develop a KV generation system capable of replicating specific styles, leveraging pre-trained LoRA models for individual products to deliver visually stunning results, even in high-resolution formats suitable for print. To build a prototype, I explored the capabilities of several AI coding tools. This article shares my initial experience and learnings, starting with the high-level PRD I used as input for these systems.

Designing KV System which is an AI system for generate Key Visual for client brands.

Tech Stack:
- Next.js (for frontend)
- Tailwind CSS (for styling)
- Shadcn UI (for UI)
- OpenAI (for prompt optimization)
- Supabase (for database)
- Python + Flask (for backend)
- ComfyUI API (for AI image generation)

Main Screen Features: (See the reference design)
- Base on the selection of KV sytle and selection of designed product, generate Key Visual for client brands.
- Once the user select the KV style, the corresponding prompt will be loaded.
- All the KV styles and prompts are stored in the database for future use.
- User can download the generated image and share it with others. Or save it to the server for future use.
- User could alter the prompt before generating the image.
- Waiting animation while the image is being generated.

Setup Page Features:
- Admin can add new KV styles and prompts to the database.
- Admin can delete the KV styles and prompts from the database.
- Admin can update the KV styles and prompts in the database.

History Page Features:
- User can view all the images he/she has generated.
- User can download the image.
- User can delete the image from the server.

Database Schema:
- KV_styles: store the KV styles, KV style image, and prompts.
- Products: store the products information, description, image, lora model path.
- KV_images: store the images generated by the users.

Top Navigation: (See the reference design)
- Logo
- Search Bar
- Main Page (right hand side)
- Setup Page (right hand side)
- History Page (right hand side)

Others:
- Desktop version only.
- Responsive design for different screen sizes.
- Error handling for user inputs.

V0: My Experience with Frontend Development

Using V0, I was able to develop individual pages quite effectively. However, the process required working on each page separately, which made it more suitable for frontend-focused tasks rather than complete application development. While I could preview individual pages and even upload images for testing, these pages were not interconnected, making it challenging to visualize the entire system as a cohesive unit.

One of V0’s strengths is its ability to generate visually appealing frontend designs after just a few prompts. The tool seems to create a Node.js project under the hood, but the backend remains inaccessible from the V0 interface, limiting its usefulness for backend development.

Once I was satisfied with the frontend preview, I could use the “download codebase” feature to export the code for integration. This feature allowed me to directly include the codebase using an npx command. However, this export process is page-specific—meaning I had to repeat the process for each individual page if my project involved multiple pages, which was time-consuming. It remains unclear whether V0 offers a way to export all pages at once.

Overall, V0 excels in generating frontend prototypes quickly but falls short in handling backend tasks or providing seamless integration for multi-page applications.

Cursor: A Promising AI Coding IDE with Limitations

Cursor has been one of the most talked-about AI coding IDEs in recent months, often regarded as a top competitor to tools like GitHub Copilot. Built on the open-source version of VSCode, it provides a familiar environment for developers, making it easy to get started. However, while Cursor shows promise, my experience with it revealed both strengths and limitations.

When I tested Cursor, I input my PRD and observed that it generated separate folders for the frontend and backend. However, these weren’t linked, requiring me to handle each part of the project independently. While Cursor could replicate the look and feel of my original layout, the generated frontend designs were relatively basic and lacked support for the latest UI components. I managed to request the use of newer components and APIs, but this led to additional challenges — Cursor often entered repetitive loops when encountering bugs and struggled to resolve them effectively.

One of Cursor’s major weaknesses is its limited capability in understanding and working with an existing codebase. It lacks robust retrieval-augmented generation (RAG), which often results in incorrect representations of directory structures. For instance, when I asked Cursor to list my current codebase directory, it frequently provided inaccurate outputs, making it difficult to locate and revise the correct files.

From my experience, the best way to use Cursor was to treat it as a co-pilot rather than expecting it to handle complex tasks independently. I relied on Chat features within Cursor to suggest specific code modifications. Cursor would explain the reasoning behind each suggestion, allowing me to review and confirm the steps before implementing changes using the provided “change” and “accept” buttons. This approach required a solid foundation in coding to validate its recommendations and handle integrations manually.

While Cursor excels at quickly generating an initial project structure, it’s essential to have a well-documented PRD from the outset to maximize its capabilities. As a tool, it’s valuable for getting started but less reliable for handling complex, multi-faceted projects or cutting-edge technologies like Next.js or the latest ChatGPT APIs.

Bolt.new: A SaaS-Based AI Coding Environment with Built-In Testing

Bolt.new offers a SaaS-based AI coding platform that allows you to develop both frontend and backend code simultaneously, with the added advantage of visualizing your project’s folder structure. One of its key strengths is that it eliminates the need for local environment setup — no more dealing with the hassle of installing packages or libraries, as everything runs smoothly in the hosted environment.

While my PRD outlined details for frontend, backend, and multiple pages, Bolt.new couldn’t build the entire project in one go. However, the platform compensates with its interactive approach — it frequently asks for confirmation before proceeding with tasks, giving the impression of a smart, semi-automated assistant that requires minimal direction. Each time a coding update is made based on my prompts, Bolt.new runs the code to test for potential issues. If problems arise, it either fixes them automatically or seeks my approval before making corrections.

That said, there are limitations. The coding speed and testing process can feel somewhat slow, which may frustrate users working on tight timelines. Additionally, Bolt.new does not allow direct image uploads for testing purposes; it only supports URL-based image inputs. While this isn’t a dealbreaker, it does limit flexibility during prototyping. I discovered that Bolt.new integrates with another IDE environment, StackBlitz, but I haven’t explored it fully. StackBlitz appears to function as a traditional SaaS IDE without Bolt.new’s AI coding support.

The platform surprised me during the final steps of my project when I asked if it could integrate Supabase to handle dynamic database data. Bolt.new generated all the necessary SQL commands and detailed instructions for configuring Supabase. While I had to manually execute these steps in the Supabase platform, the seamless guidance provided by Bolt.new made the process straightforward. In the end, I successfully connected my project to Supabase, highlighting the platform’s ability to facilitate integration with external tools. If you’re looking to prototype with minimal hassle and maximum automation, I highly recommend using Bolt.new.

Windsurf: A New Agent-Based AI Coding Platform

Windsurf is a recently launched AI coding platform, released less than two weeks ago. It is described as an agent-based coding environment designed for high automation, running your code seamlessly in a local setup. The platform’s IDE is built on the open-source version of VSCode, providing a familiar interface for developers. One standout feature is its Cascade functionality, which resembles the Composer feature in Cursor. You can describe what you need in terms of coding, and Windsurf will automatically create the necessary folders, files, and even set up all required packages for you.

During my testing, I found that Windsurf takes a highly automated approach to coding. By default, it doesn’t show you the modified code unless you explicitly choose to view it, which makes it more suited for those who want to focus on outcomes rather than details. It asks for permission before executing critical actions like package installation or code execution, and if any bugs arise, it automatically debugs and fixes them for you.

However, there are a few limitations. Windsurf currently doesn’t support drag-and-drop functionality for images or visual files. For instance, if you want the platform to adapt its code based on your custom layout or fix a preview issue, it lacks a direct way to interpret such visual inputs. While I managed to save image files to the folder directory for Windsurf to read, the user experience felt clunky and unintuitive.

What sets Windsurf apart is its remarkable accuracy in coding. According to insider information, the platform has been specifically trained on identifying coding problems and matching them with the appropriate snippets to fix them. This targeted approach results in higher accuracy and eliminates the looping issues I encountered with Cursor. Windsurf also handles the latest programming frameworks effectively, even without extensive reference documentation.

For projects where user interface design is not the top priority, Windsurf’s accuracy and automation make it a strong contender. Let’s look forward to future versions that improve visual and drag-and-drop support — enhancements that could elevate its usability even further.

Costing and Subscription Options

All four tools — Cursor, V0, Bolt.new, and Windsurf — offer a basic subscription priced at $20 per month, but there are slight variations in trial options. Windsurf provides the most generous trial period with two weeks of free usage, making it ideal for thorough evaluation. V0 and Bolt.new offer trial credits, but these are quickly consumed, especially if you are working on complex or multi-page projects. Cursor, on the other hand, has a free version with access to a limited “small model,” but this option is not recommended due to frequent bugs and reduced functionality. While the pricing is consistent, Windsurf and the limited trial credits from V0 and Bolt.new provide some opportunities to explore these platforms without upfront commitment.

Pro and Cons

Final Recommendations

• Frontend Prototyping: Use V0 for fast, visually appealing page designs or Cursor if you prefer more manual control with guidance.
• Full-Stack Prototyping: Bolt.new is your best bet for automation and platform integration with minimal setup hassle.
• Backend/Complex Coding: Choose Windsurf for highly accurate, modern programming with excellent debugging capabilities.
• Budget-Friendly Option: Windsurf’s two-week free trial makes it the most accessible for testing without immediate cost.

If you want me to show details and application of the KV system, please leave comment to me and I will unveil it for your usage.