How did I do? Mid-year reflections on my Perspectives

When you write a weekly newsletter like this one, there’s not a lot of time to pause to think about what happened last week. Once the usual Friday afternoon state of “it’s good enough, ship it!” is achieved, I start to wonder what on earth I’ll have for next week’s issue. During the following seven days, I come up with ~3k words of perspectives again and then the cycle repeats. The grind never ends.

The total word count I produce translates roughly to writing one business book every six months. Some of the material is freely available on the internet, some of it is gated behind the Plus subscription paywall. Since it’s not one big drop of book pages but rather a weekly drip of new perspectives, it can be tough for the readers to get a handle on all the topics I’ve covered and where exactly I wrote about them. Because it ain’t that easy for me either.

It sounds like the kind of scenario where AI agents would have a lot to offer. After all, what I produce is just text (plus a few meme pics) and large language models are literally built to process such data. These days, there’s an MCP for everything, and that applies to my newsletter platform too, beehiiv. Connecting a local agent like Codex to the article database is a breeze, so why not go all in with enhancing your productivity with AI?

What I’ve found useful is addressing the mundane stuff. Fixing typos and improving sentence clarity is something I’ve been asking AI assistance for a long time. What the MCP server now helps with is that I can chat with Codex / Claude Code in my Windows Terminal about which bits in the draft newsletter issue should be changed, and which should not. Then, no more copy-pasting things between windows as the agent can do the edits directly to beehiiv.

What has been less valuable to me is trying to get LLMs to surface new insights from what I’ve written. Yes, they sure can come up with lots of answers that sound very clever when I fill their context window with my organic writing. I could easily set up Microsoft Scout or whatever scheduled agent to burn some tokens on a weekly schedule and generate a personalized “Perspectives on Perspectives” newsletter just for myself. The technical part is not hard.

It all comes down to how it feels, though. Even as the models get bigger and better at some tasks, any outputs that are less about information retrieval and more about articulating the meaning of something leave a plastic taste in my mouth. On one hand, they perform an amazing imitation of a thinking machine. And at the same time, that “thinking” is not a sustainable substitute for what we feel happening in our organic brains as authentic synapses are connected.

The phrase “microplastics for the mind” is uncomfortably accurate as a metaphor for the use of generative AI in this regard. It’s fine to use plastic utensils for eating your salad in the park. It’s not okay to throw them into the bushes after you’re done. And it’s quite unsettling what the existence of plastic everywhere in the food chain eventually does to us.

Having said that, while playing around with Fable 5 last week, I did get quite nice outputs from it when pointing Claude Code at my newsletter archive. Not the deep human thought but rather a reviewer with a fresh set of artificial eyes. This way, the gaps between what’s real and what’s synthetic didn’t matter all that much, which is precisely how I think one should go about finding places to safely apply AI tools at work and in personal life.

How this worked for me was as follows: Fable 5 went through my newsletter archive and selected a set of predictions that I had made as part of the issues. Then, I sat down and reflected on how well I think those predictions held up. I graded them with [confirmed / partial / wrong / too early] labels and wrote comments, from my current self to my past self.

Then, I turned it into a web page:

Because the native website editor on my newsletter platform has some limitations, this one runs on Lovable. Again, a modern service that has an MCP server for letting AI tools like Claude and ChatGPT work with it.

This aligns with how I see the natural use cases for AI app builders in the year 2026. You’ve got your business data in a system of record somewhere, built and designed to act as a reliable, deterministic source of truth. Then, there are the approved connectors and MCP servers that expose it to allowed AI tools to process. Finally, the presentation layers are generated almost on the fly for whichever scenarios the business encounters.

In the low-code application platform era, this was often about building Power Apps canvas apps on top of tabular data sources like SharePoint lists, SQL, Dataverse tables, and publishing them internally as mobile-friendly, simplified CRUD apps. In the AI app builder era, the output can look incredibly rich, be published to any intended audience (employees, partners, customers) and it all gets “programmed” by an LLM that’s fluent with React apps.

No, I wouldn’t replace my newsletter platform with a vibe coded version. But I sure will take advantage of the new layers and tools that help me skip doing manual website building and data grouping for content that has no hard dependencies on tricky things like monetary transactions. I don’t believe the SaaSpocalypse will wipe out all of the existing systems and replace them with AI-generated code. Instead, I bet we’ll just have an ever-growing number of targeted software solutions for specific needs — instead of hoping that centralized, monolithic enterprise systems could serve the needs of all users.

Check out the State of Perspectives H1 2026 for more, and get ready for the second half with more predictions and hot takes on where business apps are heading!