The background

I was sitting in the corner with dual display setups fixing paddings, using Neovim and listening to Bhojpuri songs; still, why do people need those scroll animations on marketing websites? Someone from the sales team walks up to me and tells me that there is a new tool which can answer questions like: what is a closure in JS? Pretty impressive. It’s called ChatGPT. My first impression: they could have done better with the name.

Cut to today, my setup looks like this:

I use Claude code inside Cursor - from using Neovim to this. It’s a big change.

The goods

When you code in something which exists in the resume of every intern, every company has a codebase in this, the AI has got enough meat to train on, and naturally, will be good.

So, no surprises that GPTs are good at React, and even Shadcn, Tailwind and other popular libraries that go well with that.

• I now have to type less - the mechanical autofills, boilerplates (React has a lot of that!) and import stuff. The bottom 30% of the typing is already done by these.
• The context that it has: If I tell it that there is a function for this in the code, don’t define it again. It generally does find that and uses that.
• It is good with documentation scavenging: What is the prop which stops a Dropdown menu from the default portal renders? Believe me, it did find that.

The Bads

• When it comes to writing JS code, it generally struggles more. e.g. utility functions, taking into account the failures and other things.
• If you cannot define exactly with concrete flow, it can always mess that up - what people fail to understand is that 10 developers working on a codebase cannot follow what they like but have to follow the project guidelines. Although, I tried with the cursor rules files and all, sometimes it would deviate.
• Emitting non PR ready code: if I just check that the code is working and forward it for merge, I will get some embarrassing comments. Has happened before.
• you have to read a lot of content who you have not written - harder to reason and understand and find pitfalls.

The recommendations

What all you can delegate to the AI and have good sleep:

• The code you don’t give a f*ck about - code which feels like physics - proper rules and guardrails, not like maths where you can have multiple paths to arrive at the right answer.
• The code I have seen thousand times: I have lost the count of how many useState hooks I have called. So, probably, it’s time to delegate that now.
• The solved problems - How to add a middleware in NextJS - that’s a solved problem. Let AI replicate that.

The conclusions

• Code is not a deliverable, an asset for AI - it takes it as a medium to implement what you asked it to do and does not think that this is what you will look at for next 2 years, 8 hours a day.

• Good enough is good for AI - If it does what you asked it to do, it marks the job done.

• A job should not be the place to apply knowledge - but to learn, connect and grow. If you allow AI to do a major chunk, learning takes a back seat.

• It’s a replication tool, not a thinking machine: finds out patterns, replicates that and achieves what you asked.

• Never use while learning - it should be done banging your head, going through the docs and trying to replicate the things.

But please, start using it for day to day work. Give it a try!