To Stay Current on AI, Stop Reading and Start Building

If you want to understand where AI is headed, you need to write code. Not well. Not for others. Just write it. The most advanced AI-native workflows exist today in the tools built for software development. If you're limiting yourself right now to reworking emails in ChatGPT, you're missing a lot.

I help people with AI adoption, including non-technical people. I've spent a lot of time wondering what full AI augmentation looks like for the non-technical but I realize I've been overthinking this. Coding is the frontier. Everyone doesn't have to live on the frontier, but that's where it is. So if you want a jump start on understanding what's currently going on with tech and AI, ignore the "non-technical" label. Those lines are very blurry now and the benefits of pushing through should give outsized returns.

For myself, all of my workflows right now rely heavily upon what I knew as an engineer. I'm crafting documents and strategy, but I'm also writing code and building apps and automations. My GitHub repositories are as likely to have contracts as they are code, but the docs are definitely in markdown because that works well for my AI tools.

A Workflow That Changed How I Build

A couple weeks ago I found this blog post laying out a workflow for coding that took over my life.

Step One: Let the AI Ask You Questions First

It has a step for brainstorming where you have the AI ask for clarification and provide potential options via multiple choice. I loved this because it addresses the issue of setting AI off to do something and realizing at the end that it filled some gaps in a way you didn't expect. It had to fill them in because you didn't explore all the edge cases or communicate all the implied context in your head.

Step Two: Write the Implementation Plan Before Writing Code

After you've answered all the questions and your AI understands the objective, you don't have it start writing code. You have it write out a detailed implementation plan as if it's writing for another engineer to execute. This is set up as a skilled engineer, but one with no context on the codebase or the new feature. This clearly encourages the AI not to freelance its execution but to look at the full codebase, gather all data on how things are working and how they should work for the new feature, and lay it all out explicitly. Otherwise, some workflows can very much feel like the AI is doing improv around the edges. Everything is "yes and." It is just doing prediction of the next word and sometimes it feels like it. The explicit rigor via that framing counteracts that tendency.

Step Three: Split Architecture and Implementation Across Two AI Instances

Lastly, this workflow has you spin up another instance of your AI and point it at the plan as the implementor, then ask if it has any questions. Fascinating because it's not that you need something smarter. You are just managing an LLM's tendency to sometimes miss details. You're building a system that makes up for weakness in the components. The implementor asks clarifying questions. I pass them back to the architect. Then the implementor executes in phases interspersed with the architect reviewing the results. As I have copy paste back and forth between these two windows, it does feel like I work for the AI, but the results are hard to argue with.

This Workflow Reveals How AI Actually Works

I lay all of this out because beyond being a thoughtful approach to software engineering, it reveals a lot about how these tools work, what they are good at and what they are weak at, and how to get the best output working around those weaknesses.

The Same Workflow, Now Fully Automated

Since then, the author has automated the process via skills in Claude Code and has also ported the system to Codex. It realizes the potential to automate the grind while I participate on the plane of the what and the why. Having the architect and implementor working in tandem semi-autonomously via sub-agents in my workflow is wild and opens your mind up to all sorts of other possibilities. This shapes my understanding of where AI is headed.

Real Applications: From Voice Agents to Train Pool Scheduling

I've leveraged this workflow for client work (like voice agents) to personal projects (scheduling optimizations for my kid's carpool, which in NYC is a train pool that provides a little bit of scheduling clarity for 8 families with a lot going on). I've also been able to make strategic recommendations to some clients about how to deploy their products for software engineers with Claude Code's skills and plugin scaffolding.

Why Software Development Is the Proving Ground for AI

This works because writing code is a strong application of LLMs. Programming leans on an LLM's word-based sense making but it's also objectively provable and testable. You can write code to test code. You can build systems to scaffold around tendencies for hallucination, which allows it to be a very powerful experience even without flawless AI. Because of that, the environments for building (Claude Code, Cursor, Codex, etc.) can really lean into workflows that in other areas have to have even more human-in-the-loop dynamics.

Everything I Know Comes From Building

All of my instincts about context engineering, agents, agent orchestration, the value of tool calling, what LLMs are amazing at, what they are awful at, how to get the best outcomes. My entire point of view on what's possible today and tomorrow comes from building. I learn from smarter people, but hands-on experience is irreplaceable.

Build Something This Week

When the pace of progress slows down, these experiences will certainly be translated to more forgiving user experiences. But I've spent too much time being "aw shucks" about this and not seeing the obvious. My current thesis is if you want to be on the frontier of what's happening, first, you have to be using it daily and you have to be building. Build something this week. It doesn't matter what. A tiny automation. A Claude Code experiment. Get your hands dirty.

Lately I've been removing many of the abstractions that are built for people to engage with in favor of interfaces optimized for AI tools across both professional and personal projects. Pinky promise that's what I discuss next.