Automating Metadata Generation with Power Automate and GPT

October 2, 2024

A comprehensive guide on using Power Automate and GPT to automate metadata generation for blog posts, enhancing content management and SEO.

The text of the post goes in, and metadata comes out.

Do you want the recipe for a Power Automate flow that I used to build a metadata generator using GPT?

If you've ever looked up a recipe on the web, you know that you must wade through pages and pages of reminiscences about the author's grandparents and huckleberries and summer nights.

I see no reason to change that winning formula, although my memories may be a bit more on-topic.

In the early 2000s, CMP Media launched websites and newsletters for three print magazines -- Bank Systems & Technology, Wall Street & Technology, and Insurance & Technology.

Our jobs changed overnight.

We were now responsible not just for monthly print magazines, but also weekly newsletters and 24/7 websites.

Perhaps the biggest change was that we had to use SEO-friendly titles instead of instead of puns or metaphors.

That, in turn, made it harder for our beleaguered and patient art directors, who every month had to work with a roomful of editors coming up with different ways to illustrate abstract concepts like financial networks, payments standards, and algorithmic trading. It was much easier when they could play with a concept from the lede.

For a May 2001 article about anti-money laundering (mere months before 9/11 and the USA Patriot Act, when this became front and center), I went from New York to D.C. for an in-person interview with Senator Carl Levin (D-Mich.), then ranking minority member of the Senate Permanent Subcommittee on Investigations. Cheryl, BS&T's fantastic art director, accompanied to supervise the photo shoot in his offices.

Sen. Levin stood in his library, smiling, with his hand on a globe. I asked him if he wanted to assume a more serious look while pointing at the camera. Still smiling, knowing well his audience, he politely declined.

We used our best photo of Sen. Levin on the cover. But on the inside, the interview was relegated to a sidebar.

The story itself, about new technologies and internal controls, was illustrated with a stock photo of a $100 bill in a washing machine.

The title: "Cleaning House."

Given my lede, which compared money launderers to cockroaches, it could have been a lot worse.

MONEY LAUNDERERS using the banking system are like cockroaches infesting a New York City apartment-they've been around for centuries, they're hard to get rid of, and they scurry away when the lights are turned on.

Recent Congressional hearings have shed light on how a panoply of money launderers, embezzlers, drug traffickers, scam artists and other miscreants have taken up residence underneath the floorboards of many financial institutions. (Bank Systems & Technology, May 2001)

I asked Copilot to weigh in on the title.

The original title 'Cleaning House' is metaphorical and might not immediately convey the article's focus on anti-money laundering efforts. The recommended title 'How Banks Are Combating Money Laundering with Technology' is more direct and informative, making it clearer to the reader what to expect.

It turns out that I still tend to use metaphorical, elliptical, or even one-word titles in my posts. That's because I prefer to write for readers, not algorithms. Maybe I'll change. Maybe not.

Anyway, let's make some metadata.

Starting point: What's the theme?

After 30+ articles written for this Model Citizen Developer newsletter, I wanted to know:

1.    How would AI summarize the theme of I have written so far for the newsletter?

2.    Given that theme, what types of content should appear (and not appear) in the newsletter?

3.    If I repost this content on my own website, what metadata should I include?

For the first question, I needed a way to ask ChatGPT to give me a summary of 30+ articles. But given the limits of the context window for ChatGPT, that wasn't something I could do in a single pass.

My solution was to feed ChatGPT one post at a time and then ask Copilot to rewrite the entire theme of the blog given the new information.

I will provide a blog theme statement and a blog post.

Please respond with a new blog theme that incorporates the new post into the theme. The theme should be scoped accordingly to cover past content and the current post.

Statement: Model Citizen Developer delves into the world of low-code solutions, focusing on empowering the Power Platform community and fostering community-driven innovations.

The post follows

(content of post)

And that's how Copilot came up with a new blog theme: "Empowering Communities through Low-Code Solutions and Technological Vigilance"

I decided not to go with that theme.

Instead, I'm reorganizing my WordPress site at ivantohelpyou.com to accommodate anything and everything that doesn't fit in the Model Citizen Developer newsletter.

But if I plan to add content to a blog, it needs metadata.

The idea is that I should be able to connect related stories together to create easier entry points. That way, someone discovering my writing doesn't have to start with the latest post and work their way backwards. Instead, I can create landing pages for specific areas, with filters that pick up any related content.

I could write metadata by hand like I did 20 years ago, or even ask Copilot to write the metadata for me, one article at a time.

But I have enough content across different subject areas that it makes sense to automate this process.

The first pass is to populate the blog.

And then, I can enhance and re-run the metadata generator with different questions as I think of them. Maybe I want to pull out all the quotes, or the best quotes. Or analyze sentence length or word choices.

And get this for an idea: What if I'm helping other writers who want to analyze the metadata on their own writing?

You know, that sounds like a good business opportunity.

Well, I'm sharing the recipe anyway.

Here's the link to the recipe.

If you can figure it out on your own, have at it.

But if you're a writer who just wants to run your content through the metadata processor, let me know in the comments!