Almost Timely News: 🗞️ How To Expand and Improve Content with AI, Part 2 (2026-08-23)

Almost Timely News: 🗞️ How To Expand and Improve Content with AI, Part 2 (2026-08-23) :: View in Browser

The Big Plug

Pre-registration for my new course, AI for Writers, is now open! The course opens up Tuesday.

Content Authenticity Statement

100% of this week’s newsletter was made by me, the human. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.

Watch This Newsletter On YouTube 📺

Click here for the video 📺 version of this newsletter on YouTube »

Click here for an MP3 audio 🎧 only version »

What’s On My Mind: How To Expand and Improve Content with AI, Part 2

This week, part 2 of expanding and improving content. Last week, I talked about my general process of getting more stuff out of my head. Now that my travels are done and I have hours and hours of recordings, what do I do with them?

What I do with them is turn them into a work.

Fair warning, this week’s issue is also a very, very heavy blatant sales pitch for the AI for Writers Course from Trust Insights.

Part 1: Mise en Place

All the work we did last week of gathering the ideal customer profiles, the writing style guides, etc. are important and we’ll need those on hand. We’ll also need to calibrate against my own real writing, so I’ll have a few issues of this newsletter copied and pasted together, about 5,000 words.

I’ll be using the new AI for Writers Suite, part of the new AI for Writers course from Trust Insights (available for preorder now, USD 397); this contains my fingerprinting software to establish exactly how I write as a human, and will give AI tools the ability to measure their performance against my own.

One other thing I need to do is get writing samples from other AI models. Every AI model writes differently, especially in different families. Claude writes differently than ChatGPT, differently than Gemini, etc. and they all vary based on what data they were trained and tuned on.

The key principle to remember is to figure out which model writes MOST like you, so that you start the writing process with that model. If you start with a model that writes nothing like you, it’s that much harder to correct it. I recommend you have a model generate about 5,000 words (same as your human writing sample, try to keep it apples to apples) and do that benchmark first. What would be ideal is to take the outline of a piece you’ve already written by hand and have the AI tool write the same piece at the same length.

AI has no understanding of moderation. For example, one of Claude’s favorite things to do is bicolon rhythm. You see it most notably in things like “it’s not this, it’s that” (negative parallelism) but it’s more frequent than that. Claude’s writing has such a rigid rhythm to it that you can smell it almost solely by that bicolon frequency.

• It’s not this, it’s that

• Concept, then recap

• Intro, then concept

The challenge with AI writing is that it tends to OVERuse these constructions. They’re fine in moderation, like ghost pepper hot sauce. Even a little too much is too much, and you know the moment you have a taste.

Our counter to this is to provide it with concrete metrics and measurements of our own writing, and then force the machine to adhere to those standards. Instead of the amorphous “write like me”, we want to be specific - what percentage of my text is isocolon, bicolon, tricolon? How often do I use negative parallelism as a human, or ephiphonema? How much do I write in passive voice?

Classical AI has had these capabilities for years now to diagnose writing and fingerprint it, a field of study called stylometry. So my first step in tuning is to fingerprint my own writing, then fingerprint each AI model, pick the model that’s closest to how I write naturally, and then have that model generate the first drafts.

Here’s how I do that - I give each model the outline of what the 18 ways are topically, and have each generate their own version:

Write a 9000 word business newsletter article (+/- 100 words) about 18 ways to save on token budgets. You are forbidden to use web search tools. You MUST write from your knowledge alone. Use the attached graphic for the 18 specific ways.

It’s important to forbid the use of web search so they don’t inadvertently or intentionally copy my original language.

Part 2: Building the Scaffolding

I have all my transcripts. I have my original newsletter. I have the audience. I’ve chosen my model. My first step is to do the scaffolding. You never, ever tell AI to just go off and write something wholesale - it will fail miserably because it will try to do too much. Even with today’s smartest models and their million token context windows, the process of writing and editing from many different sources can overwhelm a model and make it forget things.

Scaffolding is a concept that comes from software development, and it makes up part 3 of the CRAFT Writing Framework from Trust Insights, architect. I have the model outline the book as a whole first, then the chapters, then write each chapter individually.

One of the keys to great AI writing is, unsurprisingly, starting with good human writing first. If you have AI generate new text from whole cloth with no human inputs, it’s going to be high probability slop. If you start with human-led original work, AI can remix it but it’ll have a lot more to work with and it’ll be novel.

Maybe next week or the week after I’ll show what a completely AI-generated edition of this newsletter would look like. I could see people being disappointed in it, but I think it would be a good exercise to show how I would do it, and specifically how I would do it in a way that was different than the way most people generate newsletters.

Part 3: Tune, Tune, Tune

Once I’ve got all the draft chapters together, it’s time to tune them. Tuning is the most important part, to remove all the weird AI writing. Part of the reason AI writes the way it does is because it has zero understanding of frequency. It’s like that novice video editor who thinks they have to use EVERY transition in Adobe Premiere or Davinci Resolve, and their work is a series of wild transitions that distract from the video rather than enable it.

The process here is to audit each chapter one by one against the original fingerprint and a deterministic set of rules to ensure that we’re removing the most egregious AI errors. Remember in part 1 how I said that AI has no understanding of moderation? This is where we impose the moderation.