The Hard Part Was Never Typing

How I actually use AI to get work done.

I’ve been using AI in my daily workflow for over a year now. Not just experimenting. Actually depending on it to get work done.

Here’s what I’ve learned: the hard part hasn’t changed. AI just made it more obvious what the hard part always was.

Most of Coding Was Never About Writing Code

This is the thing non-engineers don’t understand and engineers don’t talk about enough.

Writing code is the easy part. It always was. The syntax, the function signatures, the API calls. That’s just typing. The actual work happens before your fingers touch the keyboard.

Understanding the problem. Breaking it into pieces. Deciding on architecture. Covering edge cases. Figuring out where the complexity hides.

AI is incredibly good at the typing part. I’d estimate 80% of my code is now AI-generated. But I’m not asking it to generate entire solutions. I decide on the architecture. I break it into modules. I identify the components. Then I offload the implementation of individual pieces to the AI and review what comes back.

The thinking still takes the same amount of time. Maybe longer, because now I can actually build what I think of, so I think more carefully.

Two Weeks Instead of Six Months

Let me tell you about AccountView.

AccountView is Dutch accounting software. A client needed me to integrate with it programmatically. What I received: access credentials and a few API endpoints. What I didn’t receive: documentation, data structure, context, anything explaining how the system actually worked.

Zero guidance on how data was represented. No schema. No examples. Just raw endpoints and silence.

Here’s what made this problem nearly impossible without AI: solving it required two completely different knowledge bases. You need solid engineering skills to reverse-engineer an undocumented API. But you also need to understand how accounting actually works. Ledgers, journals, creditors, booking and releasing invoices. To make sense of what you’re looking at, you need both.

Most engineers don’t know accounting. Most accountants don’t reverse-engineer APIs. The intersection is tiny.

I started by copying the sparse documentation I had into Claude and providing as much context as I could gather. Then I queried the endpoints and fed it input-output pairs. Here’s what I send, here’s what comes back, help me understand the structure.

Then something interesting happened.

I hit an HTTP error I couldn’t resolve. The AI decided to search the web for answers. It found a blog post from 2016. Eight years old. It explained exactly how to structure the request body to pass AccountView’s validation checks.

That single piece of context unlocked everything.

Two weeks later, I had complete programmatic control over the system. Booking invoices, releasing them, attaching PDFs, working with creditors, ledgers, journals. The full integration.

Without AI? I genuinely don’t think six months would have been enough. Not because I couldn’t eventually figure it out, but because the search space was too large. Too many unknowns, too little documentation, too much domain knowledge required across too many fields.

AI compressed that search space dramatically. But it only worked because I had enough foundation in both engineering and accounting to guide the process. I wasn’t just copying and pasting responses. I was steering, validating, course-correcting. The AI amplified what I already knew. It didn’t replace the knowing.

I Stopped Taking Notes

The other major shift: I use AI as a meeting notetaker now.

This sounds minor. It’s not.

Before, I’d be in a discovery call with a customer, half-listening while furiously typing notes. Trying to capture what they said while simultaneously thinking about what to ask next. Splitting attention between documenting and understanding.

Now I just listen. I ask the right questions. I actually hear what customers are telling me.

After the call, the AI shapes the transcript into whatever format I need. Different calls require different focus. A technical discovery call needs to surface requirements, constraints, and integration points. A sales call needs to track sentiment. Which parts of the conversation got the customer excited, which made them hesitate. A project review needs to identify risks and blockers.

I use metaprompting to set this up. Before processing a transcript, I tell the AI the context: this was a technical discovery call, here are the tools the customer uses, here’s what I was trying to learn. Based on that, it generates the key questions I should have answers to and the points I should pay attention to.

Then I use that to create a structured template. Sections for Technical Requirements, Risks Identified, Open Questions, Next Steps. I run the transcript through it.

What comes out is exactly what I need for that type of call. No more reviewing 45 minutes of recording to find the one moment where they mentioned their ERP system doesn’t have an API.

The Amplification Effect

Here’s what AI actually does for me: it amplifies domain knowledge by at least an order of magnitude.

I’ve always been curious about a wide range of topics. School gave me foundations in everything from physics to economics. Years of client work added layers on top. I know a little about a lot of things.

AI lets me leverage that breadth in ways that weren’t possible before.

Right now I’m working on aerospace parts treatment, accounting integrations, wholesale order processing, and insurance claims. All at the same time. Each domain has its own terminology, its own edge cases, its own unwritten rules.

Without AI, context-switching between these would be brutal. With AI, I can go deep in each one fast. But only because I have shallow foundations to build on. The AI extends what I already know. It doesn’t create knowledge from nothing.

I’ve tried using AI in domains where I had zero foundation. It’s slower. I have to build the base first. Asking basic questions, establishing frameworks, getting oriented. Once that’s in place, everything accelerates. But without it, you’re just generating confident-sounding text you can’t validate.

This is the part people miss when they worry about AI replacing them.

The Fear Is Misplaced

People in my network keep asking if AI will replace them. I understand the fear. But they’re worried about the wrong thing.

The hardest part of knowledge work was never typing. It was thinking. Understanding problems. Asking the right questions. Knowing what good looks like.

AI handles the typing now. The thinking hasn’t gotten any easier.

I’m not worried about AI replacing me. I’m worried about other people using AI more effectively than me. That’s a different threat, and it has a different solution: get better at using AI. Find new workflows. Push the boundaries of what’s possible.

The engineers who thrive won’t be the ones who can write code from memory. They’ll be the ones who can think clearly about problems, break them into pieces AI can handle, and validate what comes back.

That’s always been the job. AI just made it obvious.