Claude Code and the State of AI Coding

Welcome to the No Longer a Nincompoop with Nofil newsletter.

Here’s the tea 🍵

• Claude Code 👾

• Cursor and AI IDEs 💻

• AI & No Code ❌

• The future of Code 🧬

This newsletter will go over the current 3 ways to code with AI. What turned out as a small section on Claude Code has turned into a newsletter covering different ways to start coding with AI.

I’ve had many people asking me how to code with AI recently.

If there’s one thing you take from this newsletter, I hope you try building something with AI. If you have ideas, AI will create them for you. It has never been a better time to build.

Now, let’s get into it.

Claude Code

Before I even get into Claude Code, I’m just letting you know that there is no more waitlist, you can try it yourself. Just run this in your terminal. You should probably read this newsletter first though.

npm install -g @anthropic-ai/claude-code

If you get a permission error, run it with “sudo” in the beginning.

Now, Claude Code (CC).

Is it good?

Yes.

Can it be amazing?

Absolutely.

Can it be absolutely terrible?

Somehow, also yes.

It’s fascinating to see it work. CC can definitely do the work of a junior dev, even a rather competent one. I would know, I was once an incompetent junior.

Unlike me, CC can, at times, be as good as a senior. It can make magic happen. It’s such a strange thing - CC can be amazing and it can also be terrible.

Let’s talk about the good.

CC is a terminal tool. It’s not a standalone application, nor is it a VS Code plugin. This means its automatically not going to be used by the majority of people on the planet, even people that have heard of it.

This is simply because most people don’t consider themselves “technical” and therefore won’t come near CC.

I think this is a mistake. If anything, if you’re reading this and you’re non-technical, I would highly recommend and strongly advise you to try CC. Not even to actually do some real work, just play around with it.

Ask it to make you a basic website or app. Just see what it can do.

Why?

Because it is so obvious that this is the future of code (to an extent). I mean, there is definitely something “magical” about telling it to implement a feature or make some changes on a website and it just goes and does it.

Once again, I would highly, highly recommend trying this, particularly if you have never coded anything before, and especially if you have.

How to use CC

Now, let’s say you’re giving it a shot. How should you use it?

Firstly, CC can read, add, remove and edit files. It can use the terminal to use git commands as well, meaning it can retrieve code from online, make changes and push any new features it adds to an online repository.

It works by using a number of different tools.

When the conversation gets too long, you can use the “/compact” command to summarise the convo and reduce its size.

For most actions, CC will ask you to approve before proceeding, although you can remove that and essentially pray it always works well.

One thing you need to understand about Claude 3.7, which powers Claude Code, is that it will never, ever, say something can’t be done. You absolutely need to keep this in mind if you’re using this for anything serious.

Claude 3.7 is not the kind of model that will try to challenge something you say - it will always try to do whatever you ask it, no matter how insane it sounds.

This is why I think non-technical people need to start using these tools.

If you know what needs to be made, you can use AI to make it. Designers who know what people want in a product are getting superpowers.

Imagine prototyping an app in 5 minutes with a few prompts.

Products like Claude Code allow you to iterate on ideas faster than ever. They let you build things you otherwise wouldn’t have. They give you agency.

The only issue is that Claude Code is expensive. Seriously, this thing will absolutely crunch tokens and your wallet.

In about a weeks worth of usage, Claude Code has digested over 72M tokens and output over half a million. I’ve barely use the console so the majority of this is CC.

It cost ~$45 USD.

Some notes on using it

Use CLAUDE.md

Claude Code creates a file called Claude.md. This file is referenced by CC to see what it has done previously. I recommend using this file as a way to keep track of progress, bugs and anything else it has been doing.

Remember, memory is one of the main problems with LLMs.

Create your own docs

If you have any documentation you want it to reference, simply create a file and put it in there and let Claude read and reference it. A quick way to do this:

• Install a chrome extension that lets you copy all contents of a web page

• Tell any other AI model to format it for a markdown file. This is necessary because there’s heaps of useless stuff on the page we don’t need

• Paste the result in a .md file in your project, perhaps put the file in a folder called docs.

• Tell Claude code to reference this file

For example, I was testing building a mobile app and added docs like this. I did the exact workflow above, taking info from a webpage and formatting it as a markdown file.

Talk to it without the code

Sometimes you want to just discuss what work is needed and you don’t necessarily want code. In this case, you have to specifically tell the model “don’t give me code” at the end of your message. An example of a message I’ve used:

“Let’s explore X idea. How can we build it? What is required? Is it feasible? Are there any considerations we need to take into account? Don’t give me code, let’s first explore and come up with a plan.”

It’s always good to ask them for options and explore ideas. This is can significantly improve the work they’ll do.

You see, LLMs think you know everything. You say “Build me a facebook marketplace clone” and the AI will genuinely try to do so. What it doesn’t “understand” is that you might not know the best thing to do. That’s why you have to prompt it to explore ideas, problems and solutions.

Remind it to keep it simple

Another thing that can help is to tell the AI to keep it simple. Sometimes, they’ll do something very simple in a very complicated way. Remind it to keep things simple.

You can add all of these things in a doc and tell it to reference this when building. I’m sure for specific industries, the AI will have different mannerisms and behaviours which will need to be guided and managed.

One very fortunate benefit I’ve had consulting for the last few years has been to see how AI does in a number of different fields. Because AI adoption is industry agnostic and I’ve consulted and built products across so many different and unrelated fields, I’ve seen how AI works in different situations.

Although I don’t think the differences are that unique, it’s the edge cases that surprise you.

Tell it to think harder

If you want the model to think for longer about a problem, simply tell it to do so.

This will make it show its thinking process.

Custom commands

You can create custom commands to quickly get things done by creating markdown files and adding them to the “.claude/commands/” folder.

Final Thoughts

The ultimate usage of Claude Code is where one instance control several others to complete tasks. This is possible and I’ve seen people do this already.

This is the ultimate usage of CC and I won’t be surprised if Anthropic release something like this later this year.

One of the engineers at Anthropic had this to say.

Lord knows what kind of systems they’re using behind the scenes with more powerful models.

UPDATE: Anthropic is planning to release Harmony, an AI agent that has access to your local files and can make edits and changes to them [Link].

If you can’t see it, AI is slowly being given the ability to do any task on your computer. Make of this what you will.

What if I don’t want to use Claude Code?

Claude Code is expensive and it sits in your terminal. I can understand why someone wouldn’t want to use it.

Moreover, what if you want to use different AI models?

Claude isn’t necessarily the best AI model for every single scenario.

This is the kind of world we have now. The landscape has completely changed. In the beginning, Claude was best for code and ChatGPT was best for most things.

Now?

Claude will solve X but fail Y and Z. GPT-4o will solve Y and fail X and Z. R1 will solve X and Z but fail Y. o3-mini will solve Z and Y and fail X.

The reality is, there is no one model for all use cases anymore.

The landscape is extremely fragmented. As models get even better, it is even harder to understand their differences.

I think this is a good breakdown for Sonnet, Grok and R1.

So, what’s the best way to try all these different models?

Code Editors

The most popular AI code editor is Cursor. It is built on VS Code which is a very popular code editor.

Cursor is definitely the best AI coding tool right now. They’re the biggest, and also the fastest growing SaaS in the history of SaaS.

And they haven’t spent a cent on marketing… [Link].

You can use it to ask questions, have it write code and use it in “Agent” mode.

Agent mode is where the AI model can simply go and make all the changes it wants. If you’re not technical, you’re probably going to turn this on and let it run wild.

So, if you’re non technical, should you start using Cursor?

It depends.

I think very recently Cursor has integrated Claude 3.7 in a better way and it’s now working very well. Absolutely anyone can build something with this.

Let’s say you want to build a website. Naturally you want it to look nice. Should you go to Cursor and describe the design of your website and pray it gets it right?

I don’t think so.

No Code tools aren’t dead… yet

Lovable is the fastest growing startup in Europe ever, hitting $4M ARR in just 4 weeks after launching [Link].

It is undoubtedly the best no code AI tool. It lets you connect to an actual database, use authentication and setup a custom domain.

The reason I think it worth trying Lovable is because in my opinion, its ability to design nice frontends is fantastic. It’s really good at designing nice websites, which is what a lot of people are looking for.

They’ve also released some new features recently which make it even easier to use:

• You can edit the actual code in a file. So you could technically give the code to Claude and update the file yourself if you wanted

• You can select anywhere on the website and target a specific button or section and have the AI edit it. This is the future of no code.

Why would you build a website using a no code tool like Webflow or Framer, when an AI can build the website and you can specify all the changes and edits it requires + set up a database and authentication, all for a fraction of the cost.

This is definitely the future of website building. No code AI native building. Tell the AI what to do, and it goes and does it.

What does this mean for the future?

So, AI is getting really good at coding, and eventually, it will be good enough for a lot of jobs also.

Anthropic’s CEO Dario Amodei has said on a number of occasions that AI will completely take over programming. He recently said that within 3-6 months AI will write 90% of the code and within 12 months, most code will be written by AI [Link].

Some thoughts on this.

Will AI start writing a majority of code within 12 months?

It’s possible. But, this doesn’t mean it will take over all of the existing coding jobs.

One scenario is that so many non technical people will generate so much code that it will be more than the amount of code that currently exists today.

This take seems plausible.

I think this is a possible scenario but, I would say it’s more unlikely than likely.

I think technical people overestimate the desire of non technical people to build things. Perhaps I’m wrong but this is the feeling I have at the moment and has been my observation also.

There is also the matter of bureaucracy.

Companies and governments aren’t just going to let AI start writing their code. This is simply not going to happen. Does this mean AI won’t be able to?

Probably not. But, this isn’t a matter of ability. It’s a matter of politics.

It’s also kind of funny that Dario is saying this while Anthropic is hiring so many engineers…

If AI can write all code, why do you need human engineers?

What’s in it for you?

I think for the average person, the questions you need to be asking yourself are:

• What can AI do for me?

• What work do I do that AI can take over?

• How can I get more of my time back by implementing AI?

• What repetitive task can possibly be automated?

In my opinion, the two biggest advantages of using AI are:

• Getting time back

• Exploring new ideas (quickly too)

An example on time - I’ve helped companies transform workflows they’ve had to go from:

• 30 minutes → 10 seconds

• 2-4 hours → 5 minutes

• 1+ day → 30 minutes

Nothing is more valuable than having more time to do the things you want.

Mind you, this never led to the companies letting employees go. It made their lives easier and they made more money.

I haven’t even gotten to MCPs yet and we’re at the end again. Next time 🫡.

Please consider supporting this newsletter or going premium. It helps me write more 🙂.

As always, Thanks for Reading.

^{Written by a human named Nofil}