One Year of Building in the AI Era: What I Know Now

The version of AI-assisted building you see at conferences is not the version you experience at 11pm on a Wednesday when the agent broke something and you have to figure out what it touched.

This is a reflection post, not a how-to. I've been building seriously with AI tools for over a year now: not using them occasionally to autocomplete text, but integrating them deeply into the way I design systems, write code, build agents, and think through problems. Here's the honest accounting.

what I was wrong about

I underestimated how much verification work the tools would create. I thought AI tools would save time in a net sense, and they do. But they also generate outputs that need to be checked, and the checking is real work. The time savings in generation are partially offset by the time requirements in verification. At the current state of the tools, the balance is still significantly positive, but it's not as lopsided as the demos suggest.

I overestimated how much the tools would change what I needed to know. I assumed that AI assistance on implementation would reduce the importance of deep technical knowledge. The opposite happened. The AI tools are most useful to people who know enough to evaluate their outputs. Shallow knowledge produces worse outcomes with AI assistance than without, because you accept wrong things confidently. Deep knowledge lets you use the tools well and catch the gaps.

I assumed the consistency would improve faster. The tools are genuinely better than they were a year ago. They're still not consistent. The same prompt on the same model produces variable quality output across runs. Building robust systems on top of them requires architecture that accounts for that variability.

what I was right about

The productivity ceiling is much higher than I thought before I used the tools seriously. I ship more now than I did before, on substantially fewer hours per shipped unit. This is real.

The bottleneck moved from implementation to specification. Writing the code is not the hard part anymore. Knowing what to build, specifying it precisely enough that the AI can implement it well, and verifying that what came out is what you intended: that's where the time goes now. This is the right shift.

AI fluency is genuinely stratifying. The builders I know who are serious about these tools are operating at a different velocity than the ones who aren't. The gap is growing. I don't say this to be smug. I say it because if you're not building that fluency, the gap is relevant to you.

what surprised me

How much the tools changed how I think, not just how I work. Writing more explicitly about what I want, specifying outcomes before implementation, verifying outputs against stated criteria: these habits are better habits regardless of whether AI is involved. The tools forced me to develop them.

How often the right answer is still "don't use AI here." There are tasks where manual implementation produces better results than AI-assisted implementation, particularly when the task involves constraints the AI doesn't understand well, or when the implementation needs to be extremely precise in ways that are hard to verify in the output. Knowing when to use the tools is as important as knowing how.

How human the loneliness of building with AI is. The tools are impressive collaborators on specific tasks. They're not the same as building with people. The isolation of building solo hasn't changed much, even with AI assistance. I wasn't expecting that.

where I am now

I'm building on a mix of things: day job that funds the space, after-hours work that builds toward something, and an honest assessment that the two need to connect more directly than they do right now.

The tools have made me more capable. They haven't solved the distribution problem, the revenue problem, or the "how do I make this work financially" problem. Those are still the hard problems, and they don't have AI solutions.

What I know now that I didn't a year ago: the tools are real, the productivity gains are real, the limitations are real, and the human questions (what to build, why, for whom, and how to survive while doing it) are not solved by any of it.

Those questions are still mine to answer.

the bottom line

A year of building with AI is long enough to get past the hype in both directions. It's genuinely useful and it genuinely changes what's possible. It also has real limitations that don't get better on the timeline of hype cycles.

The builders who are going to do well in this era are the ones who use the tools seriously, verify carefully, develop the specification skill, and keep working on the human questions that the tools don't touch.

That work is ongoing. Twelve months in, I'm still doing it.

Dru Edwards