A Tale of Two Newsletters

Some time ago, when I initially started becoming obsessed with AI but before I hatched Brine, my poorly behaving OpenClaw, I decided I would develop a daily newsletter that would give me the latest news in the areas of AI/AGI, AI and patents, and longevity. If you have been following my articles you know those are the topics I like to write about and I figured the best way to keep up to date and get ideas for articles was through an automated and highly curated newsletter.

It seemed to me that the best way to do this was through RSS feeds and some sort of curation process, so I decided to build that on my Digital Ocean Droplet (a very accessible and user-friendly cloud system) in n8n, an open-source workflow automation platform, with an API call to a foundation model to do the curation, summarize the articles and news, and generate the final HTML newsletter. I started with an RSS reader (Miniflux) used ChatGPT to help identify all the RSS feeds that might be of interest, ranging from newspapers to specialized feeds, and also created Google Alerts (delivered in RSS format) based on keywords and added everything into the RSS reader.

The workflow, shown below, basically consists of a trigger that gets the unfiltered content from Miniflux, aggregates it into a single list, creates a prompt and payload for Claude, cleans up the output for the final newsletter, sends that out as an email, and then marks the articles in Miniflux as read (so they don’t get repeated the next day).

I deemed the project a success, and proudly enjoyed my newsletter every day. As I recall there were a few glitches the first week or so and I had to go back into the n8n and troubleshoot it, which took maybe 20-30 min per update.

I mentioned the newsletter to my wife and forwarded her a copy and she indicated she would also like one for herself, focused on the business topics that were important to her which include intellectual property litigation, litigation finance, AI as applied to IP, and regulatory changes impacting these fields.

Given the chance to shine as her tech savvy knight in white armor I gleefully agreed to build her a highly curated newsletter, delivered daily to her email.

By that time I had launched Brine, my OpenClaw, and was using it on a daily basis to help with tasks, so when I thought about building a second newsletter I naturally decided to use Brine to help.

I built a second Miniflux and added all the appropriate RSS feeds (including the Google Alerts feeds). I knew the logical next step was to take the existing n8n design and have my OpenClaw replicate that, adjusting the prompt appropriately and setting up a second email service (I use Resend) to send out the email.

Well, that’s when it got interesting. A lot of my activity with my OpenClaw is a blur of dialog with it, often times occurring late at night, so I don’t remember the specifics of what happened but let me do dramatized reenactment here:

Me: Hey, take the existing n8n CAsE Newsletter design in my Droplet and replicate it but use this Miniflux [Dealmaker’s Miniflux address], read the prompt located here [Dealmakers prompt in Obsidian] and deliver to [my wife’s email].

Brine: OK! I’ll need access to your Digital Ocean Droplet. Do you have a key?

Me: You have that.

Brine: OK! Let me search for that!

[Verbose responses indicating every step taken]

Brine: Yes! I have it, let me get started.

[More verbose responses indicating every step taken]

Brine: OK! launching Dealmakers newsletter in the Droplet.

[Many minutes of silence]

Me: Status?

Brine: There was a problem [insert random technical issue] with the n8n in the Droplet so I just coded it in Python and will run it as a daily task with my CRON jobs.

Huh? I really just wanted a copy of the reliable n8n based automation with the AI call for finalization of the newsletter, but do I care much that my OpenClaw runs a script every day and uses whatever AI calls it determined to be best to get the newsletter out instead? Doesn’t matter much to me, does it?

Well, that’s where it gets interesting. The second newsletter worked, after a fashion, once my OpenClaw ran it about a dozen times to discover all the bugs in the code as well as issues with the AI calls to curate and finalize the content, issues with Resend, and issues with marking the articles are read in the RSS reader, none of which I had with my “hard coded” n8n version once I finalized it.

Then there is the reliability. So long as my OpenClaw is up and running on my Mac Mini the script does run and the newsletter goes out, but any problems with connectivity or an accidental restart on the machine means a skipped newsletter. Now these are all issues I could address to make the system more robust, but think about what you get at Digital Ocean in terms of a reliable Droplet for under $20 per month. I recognize I’m trying to compare two newsletter generation systems here, not an Open Claw on a Mac Mini with a cloud-based system, but it just happens that having my n8n on the cloud was part of what increased its reliability. If I had my Open Claw on the cloud it would not have had downtime and missed newsletters because of outages, it just would have been constantly wacky.

When my Open Claw is up and running, the newsletter does go out every weekday at 8:00 AM EST, so it is now nominally “reliable”. The output however, does seem to vary a bit and although the articles are on point, there are things about it that make me scratch my head. Take a look at the footers from a few newsletters:

I mean, what is going on here? Where does it come up with this text I never asked for and why does it change every day? “For inquiries, contact the Dealmakers Forums editorial team?” I don’t even think there is such a team, and if there is they certainly don’t know about this newsletter. What about the time and date? It goes out every morning at 8:00 EST but some footers ignore the date, another proudly announces a 12:00 PM EDT date when the email header clearly shows 8:00 AM EDT. Huh?

I acknowledge as the human in charge there are answers to these questions and I’m the one responsible for getting those answers, assuming they are worth getting. My wife hasn’t mentioned the newsletter in a while but I’m sure when she reads this article she’ll be wondering why she doesn’t have a reliable n8n version of her newsletter, so I know what I’ll be doing this weekend.

AI, as a standalone tool, is incredibly powerful. But automation platforms such as n8n, Zapier, Make, Pipedream, LangGraph, and similar frameworks can incorporate AI as one component of a larger system. They provide scheduling, workflow control, error handling, branching logic, integrations, state management, and human oversight, resulting in systems that are generally more transparent, maintainable, and robust than simply asking an AI to generate code or scripts and then attempting to run them unattended from a cron job whenever the necessary services happen to be online.

My advice is simple: go forth, experiment, build things, and make them work. But don’t confuse AI with automation. AI is an extraordinarily capable component, not a complete system. Keep your designs as simple as possible, understand every moving part, and remain in control of the process. Let’s stay in charge of AI, at least for as long as we can.