|
Almost Timely News: 🗞️ How To Expand and Improve Content with AI, Part 1 (2026-08-16)
"That was a stupid question"
|
| |||||||||||||||
Almost Timely News: 🗞️ How To Expand and Improve Content with AI, Part 1 (2026-08-16) :: View in Browser
 |
The Big Plug
Content Authenticity Statement
95% of this week’s newsletter was made by me, the human. You will see some prompt Claude wrote. 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 1
This week, I’m headed to my folks’ house to help them with some home maintenance stuff AND I’ve got a remit from Katie and the Trust Insights team to expand and improve a piece of content I did a little while back, but for Trust Insights.
I wrote a newsletter post back on May 17 about 18 different ways to save on token budgets, especially for companies doing enterprise AI. That was well received, so much so that it’s made the rounds in various companies. However, as you probably saw from that week’s issue, it was in my typical, casual style - and that’s not always as well received, nor does it capture the Trust Insights style and voice.
So since I have to be in a car for about 9 hours this weekend, here’s my process for making the most of that downtime.
Part 1: Mise en Place
Before we begin, we need to get our ingredients in order. First and foremost, I need to know who I’m writing for. The people who are going to be most concerned about token budgets at organizations will be folks like the CIO, CTO, COO, and CFO, as well as internal AI champions who don’t want an entire coop of eggs on their face when some enterprising developer burns their entire year’s worth of token budget on a two week sprint.
Thankfully, we have many of those customer profiles and personas already documented, and I can re-use that information. If I didn’t, I’d use deep research to aggregate and build those ideal customer profiles.
This part is really important because this is not who the original piece was written for, so the content will need to pivot and evolve.
We’ll also need the Trust Insights writing style as well as my AI writing humanizer (which will be part of the upcoming AI for Writers course from Trust Insights) so that it takes not only my writing that I’ve already done, but the new inputs I’ll be making.
Part 2: What I’m Missing
The original article I wrote was based on my personal experiences and my conversations with real people, real Trust Insights clients, but we don’t do business with the entire global economy. We currently have no Fortune 10 companies on our roster (call me). Nor do we have thousands of clients in hundreds of different verticals, so while I feel confident in many of the tips I gave in that original newsletter, it wasn’t as comprehensive as it could have been.
How do we change that? With the words of real people, that’s how. If I go into the subreddits for all the different AI and leadership topics, and I export the tens of thousands of posts and comments about AI specific to token usage, cost overages, etc., I will get a much bigger picture of what’s really happening in the world of AI governance on this particular topic than I could ever get on my own.
Using my Reddit developer API key and some software that I originally wrote almost 10 years ago (and has been updated many times since), I’ll go grab all those posts and put them into a notebook in NotebookLM... err, Gemini Notebook, because Google can’t stop moving the cheese.
Once I’ve got my data loaded, then I connect my notebook to Claude Code or a similar agent system and have it analyze the notebook and my original article. How? There are a ton of great tools out there; this is the MCP and CLI I use. Note I use the CLI version because MCPs are a waste of tokens.
Part 3: Filling in the Gaps
Here’s how I’ll gather the information. First, I ask what I missed - what did I cover but not in enough depth, or where my blind spots are. Claude Code called the notebook almost 60 times to ask complex questions like this:
Practitioners who tried splitting AI work into ‘plan with a big/frontier model, execute with a small/cheap model’ -- where did the small model fail to execute the plan correctly, what extra rework/QA time or cost did that failure create, and what prerequisites (like deterministic success metrics or a highly detailed spec) were missing when the approach broke down? Please attribute each claim to a specific source or thread and include short verbatim quotes with attribution where available.
It looked for evidence and came back with some real world perspectives:
MrBridgeHQ explained: ‘The drift isn’t really an intelligence thing, it’s that the plan scrolls out of the model’s attention as implementation output piles up. By the time it’s 300 lines deep into building, the plan is buried at the top of a huge context and it’s basically improvising off the last few messages.’
After that, I ask for a list of questions that are optimized for d