I did not plan to write a sequel.

But I liked my "Three Cursors walk into a codebase" article so much that, like every good movie franchise, I ultimately felt the urge to ruin it with a follow-up.

That is how this works.

You write one decent thing, people laugh, and suddenly your brain goes:

Yes. This deserves a trilogy.

So here we are. Three Claudes walk into a codebase, the sequel nobody asked for, including me.

And honestly, I am already dreaming of concluding the trilogy properly, with the final melodramatic installment titled something like:

“Brenden walks out of the codebase (because AI took his job)”

It is important to plan these things ahead. Clear road to the exit.

About a year ago, after writing about three Cursors walking into a codebase and collectively achieving 80 per cent done and 0 per cent usable, I assumed I had learned the lesson.

AI coding tools are all the same.

They generate code. They hallucinate. They stop at exactly the point where things become your problem.

Like interns, but faster.

Then I met Claude.

And just like last time, I decided names are optional.

In the previous post, everyone was called Cursor. The intern. The offshore company. The tool.

This time, everyone is called Claude.

Cursor is Claude. Claude is Claude. ChatGPT is Claude.

Only one of them is technically named Claude.

The rest just earned the title for consistency.

And now I have to publicly admit I was wrong.

We are a year further.

It was to be expected, I guess.

Claude #1 the Cursor-powered intern

Let’s start where the original trauma began.

In my previous post, I mentioned that I once asked Cursor to do a simple task I am fundamentally unqualified for: build a website with the cursed thing called CSS.

That was not a throwaway line.

That was a cry for help.

I do not do CSS. CSS is not a programming language. It is the world’s most successful hostage situation.

Claude, the Cursor, did build a website. Technically. It had structure. It rendered. It looked vaguely like something you might upload at 3 AM during a group project where nobody volunteered to do design.

Visually, it was a plucked chicken.

The skeleton was there, but the feathers were randomly glued on. Some upside down. Some are purely symbolic.

When I gave it an example, it could imitate it a bit.

But never consistently. Never cohesively. Never in a way that suggested it understood that design systems are supposed to be systems.

It was obedient.

Not creative.

Those are very different traits.

When I tried to optimise for mobile, things went downhill fast.

It would fix mobile.

Desktop broke.

It would fix desktop.

Mobile broke.

And occasionally, it would hallucinate an entirely new layout approach that broke both.

It felt like watching someone mop the floor by spreading the dirt more evenly.

I was mediating a conflict between screen sizes.

Classic Claude, the Cursor.

Eighty per cent done.

Zero per cent usable. Like a parachute that opens most of the time.

Claude #2 the OG and his team of juniors

After Claude, the Cursor, had once again achieved 80 per cent and fled the scene, I tried something more serious.

I love AWS Lambda. I do not love PHP. But some of my colleagues do, and I wanted to gently pull them into serverless without traumatising them.

So first I asked Claude, the Cursor, to Dockerise Laravel for Lambda.

He considered it.

And then confidently told me:

“Laravel doesn’t work on Lambda.”

The spirit of “not my problem” had entered the chat.

No curiosity. No experimentation. Just a polite, well-phrased version of:

Sounds hard. Good luck.

PTSD flashback to my actual intern leaving at 80 per cent all over again.

So I did what every exhausted tech lead does.

I escalated.

Enter Claude, the OG.

Claude, the OG, did not refuse. Claude started thinking.

Not typing quickly.

Thinking.

It began narrating its reasoning like an overexcited engineer iterating through container strategies without getting emotionally attached to any of them.

It sounded like a child who discovered a secret level in a video game.

It tried one approach. Then another. Adjusted Docker configs. Rethought the entry point. Reframed the runtime.

It kept going.

For thirty minutes.

I did not interrupt.

This was too entertaining. A computer was having a more interesting internal monologue than I was.

And then it clicked. It was shouting:

JACKPOT! Based on the patterns I am seeing and cross-referencing with similar attempts, this combination should work!

Laravel. On glorious AWS Lambda. Containerised. Running.

What an abomination.

I have become Dr. Frankenstein. And the monster scales horizontally with agents.

No existential crisis.No suggestion that I lower my ambitions.

It just kept pulling the thread until the sweater unravelled correctly.

Claude #3 my chatty-GPT colleague

Now, Claude, the OG, is fantastic when you want someone to disappear into a cave for thirty minutes and return with a working Dockerfile.

But he is not really a colleague.

He is not someone you casually talk to while thinking out loud.

Sometimes you do not need an engineer.

You need a sounding board.

That is where Claude, the ChatGPT, comes in.

That is the colleague who genuinely wants to listen, except they have no time to dive in deeply.

So you talk through problems together. You share snippets. You compress context and hope nothing important falls out.

As a sounding board, it's perfect. For coding, meh, it helps...until it doesn't.

But it never fully lives inside the mess with you.

Claude, the ChatGPT, is great for conversation.

Claude, the OG, is great for execution.

Claude, the Cursor, is great at reminding me that 80 per cent is still a scam.

The Real Plot Twist

You might think this is a story about picking the wrong tool. It is not. The biggest shift is not the tool.

It is me.

I stopped treating these systems like magical autocomplete engines.

Now I treat them like team members.

I document my architectural patterns. My linting rules. My scalability concerns. My security principles. My opinions about when DDD is worth it and when it is OCDDDD.

I am writing my own engineering bible.

And now my agent team members can read it in five seconds.

So instead of one confused intern named Cursor…

…I now have five juniors on steroids who absorbed my brain dump overnight.

They still need supervision.

But they learn fast.

Uncomfortably fast.

The Moral

Every developer dabbling in AI is quietly switching careers.

You are no longer just writing code.

You are training agents. You are reviewing machine output. You are setting standards.

Congratulations.

You are a manager now. The machines are not replacing you. They are promoting you.

Also, and this one hurt my ego:

Do not form strong opinions in a fast-moving field based on one bad experience.

Cursor convinced me AI coding tools were glorified autocomplete in a hoodie.

Claude forced me to reconsider.

The landscape is shifting weekly.

And if you are not careful, you will mistake a plucked chicken for the entire farm and start building your roadmap around poultry.