“By the time you sit down to write, you have done 80 percent of your thinking.”
In 2025, I first started using AI to help me write blog posts for clients. From the jump, I was excited by the prospect of turning my process into a repeatable system, specifically so that I could focus more of my energy on the really nuanced and strategic aspects of the work.
I built a custom GPT that, at the time, felt genuinely sophisticated. Its components fundamentally comprised three big moves, namely: review the brief, create a point of view on behalf of the author, and write the post.
As you may have guessed, the GPT’s drafts were underwhelming.
After delivering 46 blog posts with AI’s assistance, which involved about 100 days of writing and tinkering, I learned a lot. It became very clear, for example, that it requires way more than three steps to write a blog post that can meaningfully contribute to a high-trust sales pipeline. Because I’ve been writing blog posts for B2B firms since 2011, my process is comfortable and familiar to me. I had essentially forgotten how many micro-decisions are tucked inside each part of the workflow. To quote sociologist Annette Lareau, “[s]ome people say that by the time you sit down to write, you have done 80 percent of your thinking.” The thinking part has become intuitive to me, and I incorrectly assumed that those steps would be similarly self-evident to the technology.
“Review the brief,” for example, sounds straightforward until you realize that it involves spotting priorities, skipping bits that don’t apply, and recognizing what’s missing. A single brief might be handed off to multiple creatives and strategists, meaning that I’ll see details about graphics and visuals even if I’m hired for copywriting. In addition, many briefs lack clear proof points or fail to present them cleanly. I believe that credible writing relies on real-world constraints, tradeoffs, and specifics that reflect the actual shape of my client’s expertise; those details tend to live in background materials, internal context, or source documents that weren’t prepared with a blog post in mind, let alone a brief. In other words, my AI-assisted workflow needed a step that specifically supported discovery and data collection.
Similarly, “create a point of view” had more layers than I expected. I often ghostwrite these posts on behalf of specific subject matter experts, sometimes without having a chance to interview them about the specific topic at hand. Other times, I write pieces without any byline. While my work will always sound professional and comply with a brand voice, blog posts need some human texture to avoid feeling robotic. A little skepticism, some tempered optimism, or a hint of competitive ambition can make a piece much more readable, regardless of the specific author. I decided that crafting that subtle stance was a piece of this process that required me, the human, to be actively in the loop. That said, I wanted my GPT to source signals from potential customers, competitors, trade organizations, regulatory bodies, and others to inform my decision.
For me to better guide my GPT, I had to reexamine my process. I have since broken it into more distinct components and have seven major steps, each with sub-steps baked in.
As a result, I spend more time setting up each blog post than I used to. While it varies by deliverable, this averages to about 2.2 hours today (versus 1.9 hours). Because of this additional upfront effort, the AI-generated drafts come out dramatically closer to what I would write myself. This reduces the amount of heavy rewriting per post, and it enables more time for specific per-post tuning.
Another significant benefit of my new system is the improved resilience in my workflow. AI models regularly shift, sometimes subtly and sometimes dramatically. When my workflow was loosely defined, all I could really do was spend more time rewriting and hoping that the next prompt would somehow solve the glitch I encountered. With my more carefully articulated process, I have specific levers I can adjust.
Overall, I have been struck by how a custom GPT forces you to unpack the complexity of your own expertise. The work that you’ve been doing, perhaps unconsciously for years, cannot be transferred through osmosis. In other words, if you’re experimenting with AI-assisted writing, try to avoid chasing the perfect prompt. Instead, focus on your specific process. AI can be a powerful assistant, but only if you can clearly conceptualize, operationalize, and communicate the steps that you want it to take.
Lareau, A. (2021). Listening to people: A practical guide to interviewing, participant observation, data analysis, and writing it all up. University of Chicago Press.