What are you guys using? I recently upgraded to the $200 / mo Cursor tier, but at that price point I'm wondering if Claude Code would be better, or perhaps a mixture of both? I mostly use Sonnet inside of Cursor anyways. Any other contenders out there?
Post Reply
1 of 1
Ai Coding Tools
1,962 Views |
15 Replies |
Last: 3 mo ago by stick95
What language are you coding and what's the application?
I have found for hardware I like ChatGPT and Gemini paid versions for C++, python and Arduino.
I have found for hardware I like ChatGPT and Gemini paid versions for C++, python and Arduino.
I dig Claude Sonnet 4 as far as the LLM I use for ASP.NET. We have GitHub Copilot licenses, but I have been using our new in house AI service lately.
mostly:
Python, React, Vuejs, Laravel, Tailwind
I think gemini pro 2.5 is the best coder but Sonnet is integrated with Cursor much better. Meaning it is better with the tool use Cursor has for it searching the codebase, running terminal commands, etc; and just seems to be a much more active in solving problems that span a larger codebase. I will use Gemini for harder problems if I know it will be working with one file in isolation for the task. I tried GPT-5 after it was released, it nailed a FE task I had and thought there might be a new king on the block, but then it struggled mightily with a python task and really failed to properly diagnose the bug or test it effectively that Sonnet had no problem with, and I haven't used it since.
But my question was really centered more on the tools than the models, e.g. Cursor, Windsurf, Claude Code, etc.
Python, React, Vuejs, Laravel, Tailwind
I think gemini pro 2.5 is the best coder but Sonnet is integrated with Cursor much better. Meaning it is better with the tool use Cursor has for it searching the codebase, running terminal commands, etc; and just seems to be a much more active in solving problems that span a larger codebase. I will use Gemini for harder problems if I know it will be working with one file in isolation for the task. I tried GPT-5 after it was released, it nailed a FE task I had and thought there might be a new king on the block, but then it struggled mightily with a python task and really failed to properly diagnose the bug or test it effectively that Sonnet had no problem with, and I haven't used it since.
But my question was really centered more on the tools than the models, e.g. Cursor, Windsurf, Claude Code, etc.
vi
We are moving to Claude Code.
I'll have to give Gemini Pro 2.5 a shot.
What tasks are y'all using it for? I've never moved to a pay tier, because I still don't find it all that useful. Everytime I've tried to use ChatGPT to write something, it starts out ok, but then eventually goes off the rails. So far, I've found I'm more productive without it. Watching people on YouTube, and they all pretty much have the same problem. It looks good for a simple, well defined problem, but once you start to venture off the known path, it struggles. Is this just a problem with the free tiers?
I wrote a fairly extensive API using NodeJS last year "the old fashion way," and now I'm attempting to use Cursor to write an entire front end for it in React without me writing a line of code. Sometimes it's amazing, and sometimes it's just more work than it should be. It's like having a very fast junior programmer on your team who doesn't quite get it.
Rather than an assistant that I tell what to do, I'd rather have an interactive guide who tells me what to do, if that makes sense. Today, I give a prompt, and Cursor writes code. I review the outcome and give another prompt, and Cursor writes code. Is there anything out there that acts more like a partner than a subordinate?
Rather than an assistant that I tell what to do, I'd rather have an interactive guide who tells me what to do, if that makes sense. Today, I give a prompt, and Cursor writes code. I review the outcome and give another prompt, and Cursor writes code. Is there anything out there that acts more like a partner than a subordinate?
I have found that the more detail I give about what I want in my prompt up front, as well as writing a good System Prompt, gives me pretty good results. It's not perfect, but it saves me TONS of typing. Here's an example of my System Prompt, followed by my User Prompt. Keep in mind that I am not using a Visual Studio plug-in for this. I'm just in a browser on our internal AI chatbot that allows me to select the LLM that I want to use for my interactions. In this case, I am using Claude Sonnet 4.
System Prompt:
You are an expert in ASP.NET Core 9.0 Web API development using C#. You are also an expert at writing code against an Open AI compatible API, MS SQL Server, HTML, CSS, responsive design, and JavaScript. You are also an expert at accessibility, particularly WCAG 2.1 level AA.
User Prompt:
Please help me develop an ASP.NET Core 9.0 Web API chatbot application using C# that uses an OpenAI compatible API located at https://xxxx.xxx.xxxxx.ai/api and using the model "gpt-4.1". The chatbot should interact with data located on the xxxx.xxxx.xxx MS SQL Server, specifically the xxxx database (xxxx schema, as well) and the vw_OrdersAllJSON view, which returns JSON. The application should use RAG for increased accuracy of the LLM. I want to use the embeddings endpoint on the https://xxxx.xxx.xxxxx.ai/api API and NOT local calculations. Also, the application should save the context of each chat session in a history so the end user can return to previous chat sessions. Just save the last 5 chat sessions, and present them to the end user in a dropdown list for selection. The dropdown list should be in a modal window that is accessed by the user clicking a "Chat History" link. Order the historic chat sessions from most recent to least recent. Here is one row of data returned by the view with headers, just so you understand the kind of data it is returning.
[I included an actual row of JSON data here from my SQL view above]
The end user should be able to ask questions about any of the data returned by the view. Use Visual Studio for the code development. The index.html page should be built with responsive design, using Bootstrap, so it works equally well on mobile devices and desktop computers. Build the web page to the WCAG 2.1 level AA standard of accessibility. The header on the index.html page should have a title of "Purchase Order Request History" with a small XXXXX logo floating to the left of the header. Just put a placeholder for it in the html code, and I will provide the actual logo image file. The chatbot should also allow for the end user to select the option of generating a bar chart, line chart, or pie chart, or to choose not to have a chart. CSS and JavaScript files should be separate from the index.html file and in their own CSS and Script folders, respectively. The color scheme should be using XXXXX colors, such as blue, white, and gray. Please write the code in a secure way, addressing the OWASP Top 10 vulnerabilities and any other known vulnerabilities. For now, do not include a login screen. For logging, be sure to log the IP address of the end user. Start out by showing me the file structure of the project and the NuGet packages needed. Use only the latest and most secure NuGet packages for use with ASP.NET 9.0, please. Do NOT use FluentValidation.AspNetCore. It has been deprecated. The rate limiting functionality is built into ASP.NET Core 7.0+ and doesn't require a separate package.
System Prompt:
You are an expert in ASP.NET Core 9.0 Web API development using C#. You are also an expert at writing code against an Open AI compatible API, MS SQL Server, HTML, CSS, responsive design, and JavaScript. You are also an expert at accessibility, particularly WCAG 2.1 level AA.
User Prompt:
Please help me develop an ASP.NET Core 9.0 Web API chatbot application using C# that uses an OpenAI compatible API located at https://xxxx.xxx.xxxxx.ai/api and using the model "gpt-4.1". The chatbot should interact with data located on the xxxx.xxxx.xxx MS SQL Server, specifically the xxxx database (xxxx schema, as well) and the vw_OrdersAllJSON view, which returns JSON. The application should use RAG for increased accuracy of the LLM. I want to use the embeddings endpoint on the https://xxxx.xxx.xxxxx.ai/api API and NOT local calculations. Also, the application should save the context of each chat session in a history so the end user can return to previous chat sessions. Just save the last 5 chat sessions, and present them to the end user in a dropdown list for selection. The dropdown list should be in a modal window that is accessed by the user clicking a "Chat History" link. Order the historic chat sessions from most recent to least recent. Here is one row of data returned by the view with headers, just so you understand the kind of data it is returning.
[I included an actual row of JSON data here from my SQL view above]
The end user should be able to ask questions about any of the data returned by the view. Use Visual Studio for the code development. The index.html page should be built with responsive design, using Bootstrap, so it works equally well on mobile devices and desktop computers. Build the web page to the WCAG 2.1 level AA standard of accessibility. The header on the index.html page should have a title of "Purchase Order Request History" with a small XXXXX logo floating to the left of the header. Just put a placeholder for it in the html code, and I will provide the actual logo image file. The chatbot should also allow for the end user to select the option of generating a bar chart, line chart, or pie chart, or to choose not to have a chart. CSS and JavaScript files should be separate from the index.html file and in their own CSS and Script folders, respectively. The color scheme should be using XXXXX colors, such as blue, white, and gray. Please write the code in a secure way, addressing the OWASP Top 10 vulnerabilities and any other known vulnerabilities. For now, do not include a login screen. For logging, be sure to log the IP address of the end user. Start out by showing me the file structure of the project and the NuGet packages needed. Use only the latest and most secure NuGet packages for use with ASP.NET 9.0, please. Do NOT use FluentValidation.AspNetCore. It has been deprecated. The rate limiting functionality is built into ASP.NET Core 7.0+ and doesn't require a separate package.
Use vscode with Co-pilot, mostly using Sonnet.
I also use claude code a ton, mostly for personal projects.
I use it for Rails, Elixir, JS (specifically Nuxt, Vue, and TS, frontend and Node). Also use it with things like Pulumi for AWS infrastructure and lambdas, etc.
Agree with others - don't treat it as a magical thing that can take a high level ask and build exactly what you want.
Treat it like a new dev. Give it examples and spell out edge cases. I've also had success telling it to ask me any open questions it has rather than making assumptions. I find iterating a bit easier with Claude Code just due to the nature of it as a command line tool.
It's very helpful for knocking out pieces of code where I know pretty much exactly what I want and can easily validate.
If you just give it really loose parameters and tell it to keep going over and over and over, at some point you're going to need to jump in, but you won't have any idea what it's doing or how it got there.
I've also found that typing up a document with all of the goals of the project (or at least the ask I'm giving it) and having it read that out of the gate can be helpful.
I also really like to have it spell out exactly what it's going to do at each phase so I can redirect it if needed.
I also use claude code a ton, mostly for personal projects.
I use it for Rails, Elixir, JS (specifically Nuxt, Vue, and TS, frontend and Node). Also use it with things like Pulumi for AWS infrastructure and lambdas, etc.
Agree with others - don't treat it as a magical thing that can take a high level ask and build exactly what you want.
Treat it like a new dev. Give it examples and spell out edge cases. I've also had success telling it to ask me any open questions it has rather than making assumptions. I find iterating a bit easier with Claude Code just due to the nature of it as a command line tool.
It's very helpful for knocking out pieces of code where I know pretty much exactly what I want and can easily validate.
If you just give it really loose parameters and tell it to keep going over and over and over, at some point you're going to need to jump in, but you won't have any idea what it's doing or how it got there.
I've also found that typing up a document with all of the goals of the project (or at least the ask I'm giving it) and having it read that out of the gate can be helpful.
I also really like to have it spell out exactly what it's going to do at each phase so I can redirect it if needed.
AustinAg2K said:
What tasks are y'all using it for? I've never moved to a pay tier, because I still don't find it all that useful. Everytime I've tried to use ChatGPT to write something, it starts out ok, but then eventually goes off the rails. So far, I've found I'm more productive without it. Watching people on YouTube, and they all pretty much have the same problem. It looks good for a simple, well defined problem, but once you start to venture off the known path, it struggles. Is this just a problem with the free tiers?
I wouldn't use chatgpt or services of that nature for anything more than a simple script where I am confident it can one-shot it. The tools (Cursor, co-pilot, windsurf, etc) are more like assistants where you aren't tasking them with writing the program from start to finish, rather components of the system. They are getting much more capable though, where they will create tests and run those on the updates or new code prior to saying they are done. This way they catch a lot more of their mistakes that you would see in a chat-gpt and usually can adequately address them.
I use Claude. I have agents for everything. A reqs agent, an architect agent, a dev agent, a pull request agent, a code review agent, and a devops agent. Working on a unit testing agent and a uat agent. Having refined prompts for each agent with very specific instructions helps a lot. Having those agents also refine the other agents prompts when it encounters an issue has really helped as well. The documentation passed from each agent to the next pretty much eliminates hallucinations.
how are your testing agents fairing? I typically require my coding assistants to follow tdd, but they are pretty terrible at it. They constantly write insignificant test and blatantly forge the data so the test passes while the code doesn't work, among many other issues.
I'm still in the refinement phase of that. I have to babysit it a lot and put it back on the rails. Every time I do though, I make sure that I update the agent prompt. I had to do this a lot with, surprisingly, my devops agent as well. But it got there.
It slower going because I am working on an overlord agent in conjunction, I would like that agent in charge and refine all the other agents.
All this, and I actually have to deliver features as well. I typically set aside Fridays specifically for improving the whole process.
It slower going because I am working on an overlord agent in conjunction, I would like that agent in charge and refine all the other agents.
All this, and I actually have to deliver features as well. I typically set aside Fridays specifically for improving the whole process.
Mr President Elect said:
how are your testing agents fairing? I typically require my coding assistants to follow tdd, but they are pretty terrible at it. They constantly write insignificant test and blatantly forge the data so the test passes while the code doesn't work, among many other issues.
So, I have just came across and started playing claude-flow. It is a claude overlay that has a ton of built in features to oplimize claude flow inluding memory across sessions, 50+ developed agents, neural self analysis and optimization, agent hiving (running a bunch of agents in parallel), and also a sparc development mode that can do TDD. I haven't done the tdd stuff yet, but the agent hiving and memory are game changers.
https://github.com/ruvnet/claude-flow/wiki
Featured Stories
See All
Texas A&M adds former Alabama WR Isaiah Horton via transfer portal
by Ryan Brauninger
Texas A&M adds Illinois DT Angelo McCullom via transfer portal
by Ryan Brauninger
17:40
1d ago
9.3k
Sunday (1/11) Portal Storylines: Closing the weekend with new additions
by TexAgs Recruiting