“The first draft is just you telling yourself the story.”

— Terry Pratchett

Six months ago, I launched this blog with a simple philosophy: share field-tested experiences as “a way, not the way.” I meant every word of it. What I didn’t mention was just how close the whole thing came to never existing at all.

The Problem I Couldn’t Solve

I’ve wanted to write a blog for years. Not for ad revenue or affiliate marketing or building a personal brand — just a place to share the opinions, methods, and experiences I’ve picked up from working on real projects with real constraints. The kind of stuff you learn from doing, not from reading documentation.

The blockers were always the same two things: time and sustained focus. Researching claims to make sure I wasn’t spreading misinformation. Editing prose until it didn’t read like a hastily written internal wiki page. Fact-checking technical details that I thought I knew but wanted to verify. Each article demanded the kind of deep, uninterrupted attention that my actual work — the stuff that generates these experiences in the first place — already consumes.

So when I started experimenting with Claude Code and prompt engineering around mid-2025, I had an idea. What if I could use the same agent I was writing code with as a writing assistant? I crafted a CLAUDE.md file to encode my voice and tone, fired up an interactive Claude session, and started shaping articles collaboratively. The first two posts on this blog were born that way.

What Went Wrong the First Time

The process worked, but it was painful. I was convinced that the article needed to have just as much direct input from me as the AI writing it. In practice, this meant I was micromanaging every paragraph, rewriting suggestions, going back and forth in an endless loop of “no, that’s not quite what I mean.”

I got knee-deep into a three-part series that ate an entire week of my time. The result still didn’t sound right. I’d created something that was worse than either of us could have produced alone — too AI-polished to sound like me, too heavily edited to benefit from the AI’s consistency.

The irony wasn’t lost on me. I’d automated nothing. I’d just found a more elaborate way to spend the same amount of time and effort that had prevented me from blogging in the first place. The project quietly died.

Then the Landscape Shifted

Here’s the thing about early-to-mid 2025: the tools were genuinely in their infancy. Claude Code had launched in February 2025, and while it was impressive for the time, the agentic era was just getting started. The models were capable but limited. The workflows were clunky. I stopped experimenting for a while — not out of disappointment, but because the gap between what I wanted to do and what was practical was still too wide.

Then, over the next several months, something happened. Not a single dramatic moment, but a compounding series of improvements. Opus 4 arrived in May 2025. Then 4.1, then 4.5, then 4.6 in February 2026. Claude Code gained planning mode, background agents, custom skills. The shift wasn’t just incremental — it felt like the difference between a prototype and a product.

I started building things I couldn’t have built before. Not just bigger projects, but fundamentally different kinds of projects. My work patterns changed. I found myself reaching for the agent not as a fancy autocomplete, but as an actual collaborator that could hold context across entire codebases and reason about architecture. The 65% of developers now using AI coding tools weekly aren’t doing it because it’s trendy — they’re doing it because the capability leap in 2025 was real and measurable.

The New Approach

That capability leap eventually brought me back to this blog, and it came with a realisation: my first attempt failed because I was solving the wrong problem. I wasn’t trying to find a better way to write — I was trying to find a way to not write at all, while still pretending I was writing.

So I stopped pretending and built something honest instead.

The new approach is simple: I write a rough draft — just the core ideas, opinions, and experiences in my own unpolished words. Then I hand it to an automated pipeline that researches the claims, writes the full article in my voice, fact-checks the result, generates a report of anything that needs my attention, and even drafts social media posts. I review the output, make edits where needed, and publish when I’m satisfied.

The human — me — stays in the loop as the source of every idea and the final editor of every word that goes live. The AI handles the craft: the research verification, the prose polish, the structural coherence that turns a rough draft into something readable. Think of it less as “AI-generated content” and more as dictating to a very thorough ghostwriter who also happens to be a fact-checker.

The Elephant in the Room

I know what you might be thinking: “Great, another blog full of AI slop.”

Fair concern. “AI slop” was one of the defining content problems of 2025 — Merriam-Webster’s 2025 Word of the Year, a nine-fold increase in mentions, and AI-generated articles now making up over half of all English-language web content according to some estimates. The internet is drowning in unedited, unverified, personality-free content that was clearly spit out by a model and published without a second thought.

That’s not what this is, and I want to be upfront about why.

The difference isn’t that I’m using a fancier model or a better prompt. The difference is where the ideas come from and what happens to them before they reach you. Every article on this blog starts with my experiences, my opinions, my trade-offs from real projects. The pipeline doesn’t generate ideas — it takes mine and does the work I can’t sustainably do myself. And nothing publishes without my review.

I’m not going to pretend AI wasn’t involved. This very post went through the pipeline I’m describing. If that bothers you, I understand. But I’d rather be transparent about using AI assistance to share real experiences than either stay silent or pretend I hand-crafted every sentence.

What Hasn’t Changed

The philosophy of this blog is exactly the same as when I wrote that first welcome post. I’m still sharing “a way, not the way.” I’m still presenting trade-offs honestly. I’m still writing for people who are suspicious of anyone claiming to have the definitive answer.

What’s changed is the delivery mechanism. The bottleneck that kept me from writing — the time and focus required for research, editing, and polishing — is the problem I’m trying to solve. Whether this actually leads to more content remains to be seen. It’s still the same messy, opinionated, field-tested perspective it always was.

If anything, the hope is that this approach frees up more time for the things that actually matter: working on real projects, experimenting with new tools and techniques, and forming the opinions that become articles. The writing was never the point. The experiences behind the writing were always the point. Whether this pipeline delivers on that promise is an experiment I’m running in public.

A Way, Not The Way

I’m not suggesting everyone should hand their blog to an AI. This approach works for me because of my specific circumstances: I have things to say but limited time to say them, I care more about sharing ideas than crafting prose, and I’ve built a pipeline that keeps me in control of what gets published.

Your situation is different. Maybe you love the writing process itself. Maybe you find that the act of crafting sentences helps you think through problems. Maybe the idea of AI touching your words feels wrong in a way that’s hard to articulate but perfectly valid. All of those are legitimate reasons to do things differently.

This is one solution to one problem, tested in one person’s context. Take what’s useful, ignore what isn’t.

The ideas are mine. The words got some help. The experiences behind both are as real as they’ve ever been.