I Stopped Typing Code (Cold Turkey)

I started programming at seven years old on a Vic-20. Since then, I’ve watched abstraction layers collapse complexity again and again.

But what is happening right now feels fundamentally different.

If you work with code, data, or if you work with people who do, your tools are changing in ways you can no longer ignore. Have you tried Claude in Excel and PowerPoint?

For a long time, I treated AI coding tools as superpowered autocomplete. They were impressive, opinionated, and with errors. I regularly sparred with them pushing where they needed to go. I realized quickly that “move fast and break things” with AI can easily result in breaking small, hard-to-find things. So, I kept a very sharp eye on the output. That, unfortunately, slowed me down. But the models kept improving, the boilerplate generation got better. The scope of what I could ask for slowly kept expanding.

While trying to explain the coming impact to family and friends, I found myself thinking about the Transcontinental Railroad.

Imagine it’s early 1869. You’re traveling west from the East Coast. You take the train as far as the rails can take you, and then you transfer to a wagon to continue the journey.

But as you travel, you begin to see crews laying track from California back east, running beside the trail you had just traveled.

If you were on that wagon, you might eventually hear the work train whistle behind you before sunrise, bringing supplies to the crews laying track.

That whistle would mean something.

You might not have the words for it yet, but you would know the gap was closing. The generations before you had a grueling months-long journey. The people coming weeks behind you would cross the country in seven days.

You would recognize an unimaginable opportunity. And you might even wish you had started the journey just a few weeks later. 😉

A few months ago, I heard the whistle.

I started a new project in January 2026. Around that time, while keeping up with the steady stream of AI advances,

Given the latest lift in LLM coding capability… I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December. i.e. I really am mostly programming in English now…

I knew exactly what he meant. 80% agent coding… The interaction pattern had changed. I was getting better at providing context, sure. Preventing failures earlier, with guardrails so they didn’t repeat. Still, there was an unmistakable jump in how the tools generated the right solutions. I could aim higher. Much higher. This was no longer a theoretical improvement coming someday. The two ends of the railroad were starting to meet.

Realizing this, I made a deliberate decision.

I still write code, but I don’t type in the programming languages of the tech stack. I start with specs and clear constraints, then move to tight feedback loops, placing AI in the middle of the workflow rather than at the edges. Most of that interaction now happens in English. Often, spoken.

Recently, some non-technical parents asked what I thought about AI, the impact on education, and the future of work. Great discussion. Then came the real question: was I scared? I hesitated. We all laughed, but they weren’t wrong to ask. “Intelligence” itself is on track to become a commodity, spreading wherever it can be applied. What is the timeline? What will the impact be for us, and our children? We really don’t know what comes next.

My gut-level reaction to all of this?

Ironically, because these tools make it so easy (and cheap) to generate “correct” solutions to the wrong problems, they require more oversight, not less. Broader and broader segments of industries will need to clearly think in terms of systems. Not just those who implement technical solutions.

For the last 90 days, I’ve been stress-testing this workflow on real, revenue-impacting production systems, not toy repos. Once you do this, the hype drops away immediately, and you’re left with what is actually possible.

Here is what I’ve learned:

• Specs are your ultimate leverage.
• Defining those specs is a first-class engineering skill. Translation: it’s hard.
• Evaluation is now a first-class engineering skill. And, it, too, is hard.
• English is becoming an abstraction layer, but only when grounded in engineering discipline. (See above! Still hard.)

I don’t think the conclusion is that we no longer need makers and builders.

Instead, something more interesting may happen. People who never thought they could build systems will start building them.

I am trying to see the tracks being laid out in front of me, and I want to help others see them too before the train arrives and act early.

The expectations are shifting. Being AI-literate may soon be table stakes.

Where this experiment goes next I’m still refining this process. In my next few posts, I plan to cover:

• The challenges of moving from chat conversations to structured specs.
• The tool that made this switch much less painful (spoiler: OpenAI’s Codex App, not Claude Code for me).
• Working toward spec-driven workflows where agents run for extended periods.
• How I think about guardrails to put these tools to work safely.
• The challenges and pitfalls of incremental adoption.

Writing this post has been a challenge. I should publish it before I edit it for the umpteenth time! If you made it this far, thank you! December 2025 was the tipping point for me. I suspect there is much more to come.

Questions:

• Is this old news for you?
• Do you and your colleagues talk about this? Are you making plans?
• Are you building with AI tooling as part of your day-to-day work or exploration?
• What has changed for you in the last 3 months?

PS — Bonus Study Notes

You made it! 🎉

After I started this change, and thought I was doing a pretty good job thank-you-very-much, I saw the video below. I realized I was solidly Level 3. Well, at least it was Level 3. But the next two levels actually shocked me.

This is not just impacting developers. This… is… the point.

• Level 0 — SPICY AUTOCOMPLETE
• Level 1 — CODING INTERN
• Level 2 — JUNIOR DEVELOPER
• Level 3 — DEVELOPER
• Level 4 — ENGINEERING TEAM
• Level 5 — DARK SOFTWARE FACTORY

The piece referenced in 👆 that video: