An Update on ThePragmaticCTO
Changes on content, and what to expect moving forward
On July 21st, Substack announced its integration with PangramAI to try to combat AI generated writing or how Chris Best put "Fight against claudefishing". Now, it is no secret that I have used AI to different extents to research, write and edit this publication; when I started I wanted to experiment with AI workflows and see how close they could get to my own writing style and voice.
In 2026, there have been 4 different types of articles published on this publication: opinion and personal pieces, data-heavy trend analysis, framework/taxonomy pieces, and technical deep dives; and additionally I tried digitizing my voice and doing narrated versions of these articles.
With over 20 posts published this year, and a lot of data collected from you the wonderful readings; I can make some decisions on where to take the publication next and how to deliver the quality content that I wanted from the get go.
What is AI writing and how it as been used
Before I share more about the upcoming changes, I want to make sure I share how I been using AI as part of my workflow and provide reassurance. Let me get this out the way while I used AI to assist me on the writing of the articles; I never delegated the thinking or the ideas to AI; what see here is what I consider interesting or relevant at the time.
In fact, to share a little insider baseball my process so far has been a 7 stage workflow that goes all the way from topic research, refinement, outlining and drafting. I went as far as training the agent to have a "Voice DNA" in other works a fingerprint of how I write.
Checking flow and cadence, checking grammar, punctuation, generating figures and quotes are all things that I delegated to the AI. There has always been human review, editing, writing in the process.
Over the past 6 months, I had different levels of success and looking back; the process has a major flaw when it comes down to specific post archetypes:
Data heavy posts
Trend analysis
Regardless of how much rewriting, clean up and reprompting the flow, cadence and some of the LLMs fingerprints are still present. It is not simply the use of verbal cliches like actually or "Not X, Its Y" or other tells from LLMs; fundamentally the flow of the posts would be polluted by the mechanical touch of an LLM.
What the readers think
The numbers of also validate this, you as readers engaged much differently with the opinion pieces that where clearly in my voice, and didn't had the same creeping issue with LLM flow.
In fact, the top 8 pieces in this substack are purely opinion and personal pieces; while bottom 6 are just heavy data or trends analysis, now while you argue is the content type and not writing workflow; I honestly believe that in this day and age people don't want to read slop no matter how well polish it is.
Simplifying
Based on the performance, the cost and effort to write them I've decided to park data heavy articles and trend analysis pieces for now, and instead focus purely on opinion, personal pieces and similar content that is more natural to my writing and thinking style.
There might be new content types coming soon, and depending on feedback I might revive trend analysis. You let me know what you want in the comments
What to expect moving forward
Alright, so what you expect moving forward from this publication. Well, to start I'm drastically changing the workflow to avoid the AI assisted component to bleed into the finished piece.
I will still use AI to help me with research of a given topic, generating the mechnical parts of a post like figures, hero images, etc. But when it comes down the meat and potatoes of the article, I'm putting very strict guardrails on what I will allow LLMs to touch or change to mention a few:
No more LLM corrected grammar and type corrections, replacing with Vale and manual editing
While the LLM will research the topic, how I ingest that and use for the writing piece changes to make sure there is no bleeding of LLM phasing. More factual and reference than summaries or generate briefs
Tagging of what was actually AI generated on every post like figures, and footnotes
No more voice correction from the LLM, one of the reasons I did the Voice DNA is to try to enforce a consistent tone of voice from this publication. I'm removing that from the workflow
So as result, you'll see more types, weirder grammatical structures, and every now and then a more inconsistent voice; personally I think this overall a positive change and will make the following pieces a bit more interesting; and to be honest a bit more spicy, as I can be extremely opinionated and part the AI was to moderate that aspect.
Let me know what you think in the comments below.