The Future of My Work with AI: A Narrative Overview

A few years ago my work was mostly my hands. I wrote the articles. I picked the products. I formatted the pages, set the links, checked the tracking. If something needed doing, I did it. The output of the business was capped by how many hours I could stay at the keyboard.

That isn't true anymore. And the shift wasn't a productivity tweak — it changed what "my work" even means. This is the long version of that story, and where I think it's heading.

Act One: how I got here

To explain the now, I have to go back. Long before AI was the thing every founder talked about, I was an early employee at a crypto-fintech startup — employee number four, hired by one of the co-founders to jump-start marketing while the whitepaper was still a draft. I was the first marketing lead. We went from a few dozen people to several hundred, and the platform grew to close to two million users.

My job there wasn't writing. It was systems. I was the solution architect for the marketing technology stack — custom audiences, segmentation, email automation journeys, paid media across the big ad platforms, attribution. I built drip campaigns for the corporate, institutional, and lending teams, wired CRM pipelines that auto-assigned leads to the right business-dev person, and ran experiments — A/B tests, targeted promo campaigns, localization, custom segments by region and user behavior.

Then the company hit a restructuring bankruptcy and cut most of the staff. I was one of the people who got laid off. Years of building, gone in a memo. That ending mattered, because it pushed me back to my own projects — and that's where the AI story actually starts.

Getting laid off forced the question: if I'm building for myself now, how much of this can run without me?

The pain, first

I went back to running my own affiliate sites. Same instinct as always — do the work myself. Research a product, write the page, localize it for another market, publish, track. One page at a time. It worked, but it was slow, and slow doesn't scale across markets. I'd spend a long stretch of a day producing a handful of pages and call it progress.

The turning point

Then I started using AI seriously — not as a novelty, as labor. The first time I handed a draft to a model and got back something I only had to edit instead of write from scratch, the math changed. Work that used to eat most of a morning collapsed to a fraction of it. And the part that mattered: it scaled. Doing it once or doing it a hundred times cost roughly the same effort on my end.

This wasn't optimization. It was a different shape of work entirely.

From template to system

Each round, I removed myself a little more. First I templated the writing. Then the localization. Then the research that fed both. Within a few months I could push hundreds of articles across dozens of countries without touching a keyboard for most of it. I wasn't the writer anymore. I was the person who designed the thing that wrote.

Somewhere in there, the bigger shift happened. I stopped using AI as a faster writer and started using it to hand off whole categories of decisions — which products to feature, how to phrase titles for different markets, which calls-to-action pulled their weight, when content had gone stale. I'd moved from content creator to system architect. The work changed. So did where my time was actually worth something.

Act Two: what my work actually is now

Zoom out and my work today splits into three layers.

Layer one — system design

The strategic layer. What problem am I solving? Which markets matter? Which partners are worth the effort? What does it cost to acquire attention versus what it returns? This is pure thinking — reading data, talking to people, testing hunches. No AI does this for me yet. It's roughly a third of my week, and it's the third that matters most.

Layer two — building the templates

The architectural layer. How do I encode the decisions from layer one into reusable prompts and skills? What changes per use, what stays fixed, what guardrails keep the output from going sideways? AI helps here — it helps me refine prompts, find edge cases, trim cost — but I'm still driving. This used to be the biggest chunk of my time. It keeps shrinking as the templates mature.

Layer three — exceptions and optimization

The operational layer. Did the system do what I expected? Is tracking firing correctly across every market variant? Does it need adjusting for a season or a sale? What broke, and why? This is review, measurement, and debugging. AI flags the anomalies; I make the call. Also shrinking as automation gets better.

Act Three: where this is heading

I don't know the future. But the direction feels clear enough to bet on. Four things I expect.

Properties that update themselves

Promo codes, prices, and offers change constantly. Today I still nudge those updates along. Before too long I expect my sites to watch the sources themselves, regenerate the affected pages, test the new copy, and only ping me if conversions dip or something errors out. I go from "publish weekly" to "check dashboards weekly." Creator becomes curator.

One system across many properties

Right now I run several sites mostly as separate things. The next phase is one system orchestrating all of them — a shared skill ecosystem for content, one dashboard across properties, one loop deciding where effort should go. I'm not doubling down on a single business. I'm building a meta-system that optimizes the whole portfolio. My job becomes allocating attention and capital, not operating each site.

AI as partners with constraints

I've stopped thinking of my AI tools as tools and started thinking of them as partners with clear mandates and clear limits. The model is my research partner — it can read a pile of sources and pull the patterns, but it can't pick the strategy. The automation layer is my operations partner — it runs on schedule and escalates exceptions, but it can't prioritize between competing fires. The edge layer is my deployment partner — it serves content everywhere, but it can't decide what content matters. Standards like MCP make this cleaner every month — agents plugging into tools through one shared protocol instead of glue code. I sit at the intersection, making the calls that need judgment. That's more scalable than "use AI for everything" because it respects what AI is actually good at.

The skill gap is closing

Here's the uncomfortable part. The skills I leaned on — SEO, copywriting, conversion tuning — are getting commoditized fast. Anyone with an API key can generate a decent landing page. Anyone with a workflow tool can automate publishing. The advantage is moving away from doing the work.

The skill isn't writing one good article. It's knowing which thousand articles to write, and why.

The moat isn't tools. It's judgment and architecture. The people who win next aren't the best writers — they're the ones who became system architects while everyone else was still racing to write faster.

Act Four: the skills worth catching early

If you're building something solo, you're choosing between optimizing for the skills that pay today and building the ones that'll matter soon. Here are the five I'd bet on, and how I picked them up.

1. Systems thinking

The foundation. Not about AI — about seeing the template under a manual task, and telling apart the decisions that are rules (automatable) from the ones that need a human. If you can do this, you can adapt to any tool. If you can't, you'll chase tools forever.

I learned it by documenting every decision inside a reusable skill file, by maintaining content across many markets until it was obvious which parts scaled and which didn't, and by migrating my sites across a few different stacks — each migration teaching me what the system actually needed versus what was just nice to have. To start: pick one repetitive task, write down every decision you make doing it, turn those into rules, and see if a model can follow them.

2. Data literacy

You can't optimize what you don't measure, and you can't measure what you don't understand. Can you read a dashboard and spot what's off? Tell correlation from causation? Design an experiment that answers a real question? Read the story the numbers are telling instead of the one you wish they told?

This is what separates people who use AI well from people it fools. My systems produce a lot of pages. Are they all good? No. Some convert poorly and some convert well, and without measuring I'd never know which to clone and which to kill. I learned it the slow way — tracking everything, building scrappy dashboards, and running a long string of tests where I was sure I'd win and lost about half. Humbling, educational.

3. Prompt architecture

Prompt engineering is becoming table stakes. Prompt architecture is the next rung. Engineering asks "how do I word this to get a better answer?" Architecture asks "how do I structure prompts so they're maintainable, reusable, testable, and cost-efficient across a hundred different uses?"

A handwritten prompt works for one article. A proper skill file works for thousands. The things to learn: variables and templating, breaking complex tasks into steps, output formatting your system can consume, cost control, and failure modes — what happens when the input gets weird, and how you guard against it. I learned by building and then maintaining a stack of skill files for months, which is what reveals the difference between fragile and robust.

4. API thinking

Most people use AI through a chat window. Fine. But to build autonomous systems you have to think in APIs — rate limits, cost per call, latency and caching, what happens when a request fails, and chaining calls together. That's the jump from "AI as a tool I use" to "AI as infrastructure I build on."

It matters for unit economics. When I first scaled content, I naively made one call per article. At volume, the cost adds up fast. Once I added caching and batching, the cost per article dropped by more than half — and suddenly the math worked at a much bigger scale. To start: build one simple integration, measure what it costs, add caching, measure again, and understand exactly what changed.

5. Conviction in uncertainty

This one isn't technical. As AI eats the execution layer, the value of human judgment goes up — and judgment needs conviction. The ability to gather incomplete information, make a call, build a small test, learn, and adapt.

I can't know a new market will work across dozens of countries. I build it anyway and let the test tell me. I can't guarantee the system I design today still holds a year out. I build it anyway. I practice this by writing down my decisions and the reasoning behind them, reviewing them later to see what surprised me, and sharing the reasoning — not just the result — so it can be judged on more than luck. This is the skill AI can't automate, and it's the one getting more valuable as everything else gets cheaper.

Act Five: what I'm building next

Three threads are converging. Properties that maintain and test themselves while I check in weekly. A portfolio managed as one system that moves resources to wherever they matter most. And writing — like this — that documents how the systems get built, so the skills are catchable before they're obvious.

The honest framing: the work I'm doing isn't really about affiliate marketing. That's just the test case. It's about building the infrastructure for content and business operations that aren't bounded by one person's execution time. The same principles apply to product docs across languages, to thousands of product descriptions across regions, to newsletters tailored by segment — any knowledge work capped by human hours.

The infrastructure is the asset. The specific application is just where I'm testing it.

Why this matters if you're building alone

Solo, you're up against teams with bigger budgets and audiences. Your edge was never resources. It's leverage — and leverage increasingly comes from systems that run without you. Naval's line keeps proving true: play long-term games with people who compound, and let permissionless leverage do the rest. I'm building in public so the compounding is visible, not just claimed.

The skills I'm chasing — systems thinking, data literacy, prompt architecture, API thinking, conviction in uncertainty — are what let one person build something far bigger without hiring proportionally more people. The hard part isn't the technology. It's the mindset move: from "how do I do this faster?" to "how do I build a system that does this without me?"

What I'll do next

Keep removing myself from the execution, keep moving up into the judgment, and keep writing the steps down as I go. If the bet is right, the advantage of the next few years is architectural, not tactical — and the people who notice early get a head start. That's the future of my work. I think it's the future of yours too.