The faster I moved, the further behind I got (True Story)

A few months back, I was building a small app.

I'm not a coder. But Claude made it feel possible — I put in a rough idea, and within hours something real was running on my screen. I was genuinely excited.

So I kept going. Feature after feature. Day after day. The app was taking shape faster than anything I'd built before.

Around day three or four, Claude flagged something. A database relationship that wasn't set up right. It suggested a quick fix. I said yes. We moved on.

Five days in, I hit a wall. Not a small bug. A foundation problem. The database structure I had never properly thought through — never stopped to explain clearly to Claude — had been quietly causing problems from the start. The quick fix had hidden it. Now it was everywhere.

I had to stop. Go back to the beginning. Rebuild what I had skipped in the excitement of moving fast.

Sound familiar?

There's a difference between output and progress

Here's what I've come to understand after that experience — and after making the same mistake with writing, with building prompts, with almost every new thing I've tried with AI.

Speed with AI is intoxicating. You put something in. Something impressive comes out. Your brain reads that as progress. So you keep going.

But output is what the screen shows you.

Progress is what actually builds over time.

When I started writing blogs with AI, I could produce something faster. But the early outputs lacked context. They lacked my voice. They didn't sound like me talking to a manager — they sounded like an article. I was producing faster. I wasn't getting better.

There's a word for what I was doing: moving. There's a different word for what I needed to be doing: compounding.

What the world is learning the hard way

Klarna is a good example.

They replaced 700 customer service employees with AI. The results looked impressive in month one — faster responses, more handled, lower costs. They celebrated the speed.

A year later, they started rehiring. Not because AI failed completely. But because the complex cases — the ones that actually mattered — were falling apart. AI had given them speed. It hadn't given them accuracy on the cases where accuracy was everything.

The same thing showed up in code. GitClear, a developer analytics firm, analysed 153 million lines of AI-generated code and found something uncomfortable: developers were producing code significantly faster, but long-term code quality was quietly dropping.

They called it "AI-induced tech debt" — problems that don't show up on day one, but build underneath until one day you hit a wall.

More output. Less compounding value.

That was my database problem. At scale.

A fair pushback

Someone could reasonably say: you have to move fast first. You need the reps. You can't develop good judgment without first using the tool a lot.

And honestly, they're right — up to a point.

When you're just starting out with AI, experiment fast. Try things. Break things. That phase should be low-stakes and quick. That's how you build familiarity.

But there's a shift that most people miss.

When you move from experimenting to actually building real value — in your work, your decisions, your outputs — the rules change. Every shortcut you take at that stage gets built into your habit. And habits, as I learned with that app, are expensive to fix later.

The reps matter. But reps without reflection just build the wrong habits faster.

What actually compounds

After the app experience, something shifted in how I approached AI.

I stopped chasing what it could produce and started paying attention to why it was getting things wrong.

With the blogs, I slowed down. I built a project with proper instructions. I gave Claude context about my audience, my frameworks, the way I think and write. I wrote out what good looked like — and what didn't. I started refining the setup instead of just rewriting the output.

It took longer upfront. But after a few weeks, something changed. The outputs started sounding like me. The gaps got smaller. Each session built on the last.

That's when I understood what compounding actually feels like — not faster output, but output that keeps getting better with the same effort.

What this means for you

If you've already started using ChatGPT or Claude — good. Keep going.

But pick one task you use it for regularly. Honestly ask yourself: am I just producing faster, or am I actually getting better at this?

If the answer is just faster — that's your signal.

Here's one thing to try this week:

Take 20 minutes. Build a simple instruction set for that one task.

Write down your audience, your context, what good output looks like, what you're trying to achieve. Put it inside a Project in ChatGPT or Claude — not a new chat every time. A persistent setup that remembers.

Do it once. See what happens over the next two weeks.

You're not looking for a perfect system on day one. You're looking for the first layer of something that compounds.

That's the shift — from AI that works, to AI that keeps getting better at working for you.

Stay grounded. Stay curious.

-----

This is part of my ongoing work helping managers build an AI mindset—practical ways to think, decide, and work in the age of AI. It's not about mastering tools, but understanding how to work intelligently with intelligence. Looking forward to your feedbacks