This blog was featured on my LinkedIn, copying it here too 🙂

This started as a simple status update about the fun I’ve been having with Claude and it turned into this… strap in…

I’ve been messing around with LLM/AI stuff for a while now… personal productivity stuff, just seeing where it fits day-to-day, both at work and at home. I haven’t drunk the Kool-Aid, as the Americans say, but I’ve been cautiously following along.

I’ve used ChatGPT (personal) and Copilot (professional) loads, chatbot style. Everything from helping me structure my thoughts when writing a statement of works after a presales chat, to using it as a sounding board when I’m doing some semi-risky DIY and I’m about to do something I’ll regret. Sometimes it feels like “Google on steroids”. Sometimes it feels like a really eager junior who really wants to help… and sometimes gets it wrong with full confidence.

So far, useful… but not “wow”.

That changed when I started dabbling with Claude Code.

I confess I am a T-shaped consultant. I have good and broad knowledge of most areas of my technical sphere, and have deep knowledge on one or two areas – classically T-shaped, exactly where I should be – ducking, diving and getting stuff done. One area, in particular, which is seeing massive growth is data engineering with Apache Spark and Python (PySpark/SparkSQL). This is one of my shallower areas. I have always favoured Power BI, Qlik and SQL as my deeper areas, however you move where the market takes you.

Claude Code has been completely remarkable for upskilling in data engineering with Spark. There is something about the command line, and the way Claude interacts, talks, questions, pokes, prods. It feels like entering the rosebud cheat code when playing The Sims (or R1, R2, L1, R2, Left, Down, Right, Up, Left, Down, Right, Up for the GTA fans).

It’s not magic… it’s because the feedback loop is incredible.

This afternoon I had a technical challenge to change an API extract from a truncate and load into incremental but with a few edge cases that needed to be considered (what if records are deleted in the source system, what happens if updates change IDs, how do I make sure I can re-run without losing data – blah blah, boring stuff right?).

Normally, I’d open DevOps, write a little plan, decompose the tasks, spend half a day scratching my head bouncing between the IDE and the plan, getting frustrated and drinking tea, tinkering, coding, testing.

This time I thought, lets try this Claude Code thing out.

So I started a new Claude session, explained the challenge and my rough plan, and Claude asked me if I wanted to go into plan mode – aye ok, I will. This is where the Claude good stuff starts. Instead of going off and just “vibe coding” a load of slop (e.g. type prompt, copy code, run, hope it works, repeat), it actually slows you down in the right places, gets you to think, let the marbles roll around upstairs.

Anyway, back to me and Claude… we got into a fairly human conversation about the route forward, the upsides, downsides and trade-offs. A plan was formed, documentation was written, tests were designed, and if you think about it, a proper specification was developed. Then Claude asked… shall I implement these changes?

I am not a developer, my programming skillset is from an application support background, low-code rather than a hard-core pro-coder. The code it wrote, was, good. It wasn’t just like good enough for a demo. It was readable, it made sense, it was very good, and it worked, first time.

I added a few tweaks, made sure I understood what it’d done, checked the outputs properly, and went with it.

So aye. This feels crazy.

Also feels scary.

And a bit dangerous.

Because I can’t help thinking about the amount of Excel Hell and VBA spaghetti monsters already running businesses. Now we’re basically giving everyone the ability to generate more logic, faster, with more confidence… and not necessarily more understanding.

The risk isn’t “AI writes bad code”. The risk is bad code that looks good enough. Unforeseen data bugs. Weird choices. Stuff that “works” but is wrong. Half the battle is knowing what to ask it, and the other half is knowing what not to trust.

So, what’s next?

Well, I worry about the future.

I worry about juniors and people starting their careers. How will they start their careers when the robot can do those types of roles?

I worry about seniors too, we’re about to have the Power of Greyskull in our hands, and the ability to cock-up at scale.

I think I’ll benefit… I’ve been in tech for decades. I learned today that the debut Arctic Monkeys album came out 20 years ago. I was in college then. However, I’ve 10x’d my skillset overnight. How can you x10 when you don’t know the basics, when you aren’t battle-scarred of running a fleet of MS Access databases powering a month end or coding by the seat of your pants, propping up green-screen progress-based apps? Crucially, I have also just given myself the ability to x10 my mistakes too – the ol’ dunning-kruger effect could be in full swing.

I’m not sure how this will end yet. I’m only talking about my personal experience over the past few days. I’m optimistic about what skilled people will build with this, and how fast we’ll develop our skillsets and be better at what we do. But I also think we’re about to drown in slop… and down in people trying to sell the dream. The winners won’t be the people who can prompt the hardest. It’ll be the people who can review, who can plan and who can ship good stuff responsibly

I’m excited… but cautiously so, because this could get out of hand rather quickly. It’s going to be fun though!