Home of Wituz

Coding with LLMs for productivity and fun

AI development tools are currently booming. Tools like Claude Sonnet, Copilot Agent Mode, Claude Code, Cursor, Windsurf, Cline and so on - are awesome. But can also be overwhelming, given the amount of choice.
Some tools are good value for money, while others throw it out the window. Some of them work really well for some scenarios, while others don't really perform at all.

In this guide, I will attempt to share my experiences, so you can hopefully get inspired for your own AI workflow.

Expectations

If you are not familiar with AI coding, this will change the way you look at code.
For me, I went from manually typing hundreds of lines every day, to making the same thing possible in a few minutes, just by describing how I wanted it done.
So much so, that meanwhile I am writing this guide, an LLM is working on my project as I am writing this. For once, I do something productive meanwhile, other than leveling up Mining in OldSchool Runescape.

I am going to review each tool I use, and then at the end, I will write how I use them together.
Let's start with Copilot Agent mode.

Copilot Agent Mode

To set up Copilot Agent Mode, download VSCode Insiders and install the Copilot extension. Then you can press CTRL + Shift + P and seach for Copilot Edits.
A window will open. In the bottom right, you can select mode. By default it says Edit, but then you can set it to Agent. You can also select a model for Agent mode.

Copilot has been a neat tool, tailored for VSCode and many other (primarily Microsoft).
Until a few weeks ago, Copilot could only do editing in the file you were currently in, or give inline suggestions.
This has now changed, since Agent mode has released, which is aware of your entire workspace and can search, read and edit files across it.

Agent Mode has several options as of now for which model you want to use.

My experience with the OpenAI models in Agent Mode is that it does it well, but is generally slow and does not reflect a lot on what it is doing, leading to code that is not what you wanted, or changes you did not want.
Where the OpenAI models shine in Copilot are in Edit Mode or fixing inline issues. When Claude fails for those, I switch to an OpenAI model such as o3-mini, and I have the issue solved in usually a few seconds.

Claude Sonnet with Agent Mode is what I use the most. While Claude Sonnet 3.7 performs the best, and delivers really amazing code, the rate limit is not so generous, and you only get a few chats with it before it says you have to wait for an hour or so. I guess also as of now, the demand is extremely high for this model, and before a better deal has been made between Microsoft and Anthropic, I can understand that they rate limit it to not loose too much money.
When that is said, Claude Sonnet 3.5 has an amazingly high rate limit. I can use it for hours, and even intensely, before I get hit by a rate limit.
That's why my primary way to use Agent Mode is Claude Sonnet 3.5.
You get more for your money than with an API key directly to Anthropic, because Microsoft in this case pays for your usage.
With my API key to Anthropic, I can easily use $50 in a day, but through Copilot which is just $20 a month, I get so much more value for money.
Once I hit the free tier, I switch to the Anthropic API, and then go back to Copilot Agent Mode once it is restored from the rate limit.

Occationally you run into a bug.

Claude Code

To set up Claude Code, register for an account at Anthropic, fill $10 in it and get your API key. Then download Claude Code.

Claude Code in my opinion delivers the best code companion in the world, as of now.
The code quality is amazing, and it actually gets the job done, even for very complex tasks.
The only issue is the pricing, as it easily can burn $100 during a productive day.
If I had all the money in the world, I would use this tool.
But you get it through Copilot with a few bugs and more or less unlimited use.
That's why I primarily use VSCode with Copilot.
But if I hit a wall, I sometimes try with Claude Code, and it always works.

Cline

To set up Cline, simply install the extension in VSCode

Cline is an extension for VSCode and other editors as well.
It is just a gateway to use a similar feature to Agent Mode but with any API gateway you might want to use.
You can even use local models, but for that I prefer to use Continue as it supports LLM Studio - not something I do that often anyways, just to try out a model. For instance when DeepSeek released, I tried it through Continue.
I find that for Claude Sonnet, it is just as expensive as Claude Code since it uses the API directly, but does not always deliver as well.

Zed

A hidden gem I think is Zed.
Zed is an editor just like Cursor, by the creators of Atom.
Zed does not have Agent Mode yet, unless you add all the files it should edit manually.
What you do get though is free API access to Claude Sonnet!
Zed is a super nice IDE, and is for once not a fork of VSCode. This means it performs really, really well, and is much more snappy and friendly to your PC than VSCode.
Once it gets better AI tooling, I will switch to it as my primary IDE, but for now I stick with VSCode.

Cursor

This is the alternative to Copilot + VSCode. It works really, really well - but it does not come without issues. Often I run into issues where it suddenly just stops generating code, or throws 400 errors at the server.
I think Cursor is good if you don't mind the errors, but for now I think VSCode is more worth the money. With cursor you also get access to Claude Sonnet unlimited, but you have to wait in a queue if you use it a lot.

MCPs

A very hot topic at the moment is MCPs. An MCP can be hosted on a server, and you can then link that server to your LLM.
An MCP is an extension to your LLM. It allows it to use service it otherwise would not have access to.
An example of this is Stripe. Stripe has an official MCP that you can plugin into Cursor or Claude Sonnet, and then the LLM will have full capabilities to your Stripe account. That means you would not have to set up any products, API keys, development environments and so on, because suddenly your LLM is context aware.
There are many incredible MCPs, and I encourage you to try out a lot, as many of them are just amazing addons.

My workflow

My primary tool is Copilot Agent Mode with Claude Sonnet 3.5. Primarily because it is good value for money.
If I hit a wall, or have something very complex, I use Claude Code. It gets the job done, but costs a lot.
I hope Zed will get better at Agent Mode, if so I will switch to it, since they have full free access to Claude Sonnet.

Some tips if you are new:
- Describe your changes as precisely as you can to the LLM. The more abstract a prompt you give it, the less the chance is that it works well. Use your programming skills to supply it. Tell it exactly what it needs to know, which framework to use, to generate a class, interface, model, and so on. I usually go from somewhat abstract to more and more precise, if my prompting fails.
- If it ends up generating something that leads to an error, just tell it about the error. It usually fixes itself, and sometimes it is just unable to see what the console error will be. Often times, it will also insert console logs in your code, which you then paste into the LLM and from there it knows what is wrong and can fix the issue.
- Many people give up and think "it's not good enough" once they see the LLM make a crucial error. Once I decided to take the responsibility of how my LLM generated code, and saw it as a skill, my productivity skyrocketed and I only became better and better. I encourage you to think of these tools as skills to master, just llike you would when learning a new programming language or an instrument.

I hope this article inspired your coding with LLMs.
Please like the article if it helped you, and I would love to hear about your workflow in the comments.