Figuring out how to run learning loops

Using AI
1

We just built Radar, an AI product and design news publication that watches 60+ sources, groups the signal into topics, and publishes daily with about 95% running on its own. Two weeks, start to finish, with our side time to figure out what works. We shared five takeaways in our LinkedIn post.

image
image1536×1024 505 KB

Moving from skills to sequences

The big learning is how we’re starting to think in loops. Building our Glare Skills was the first step toward helping teams learn faster (more work to do here), but creating sequences that run regularly is the next step to unlocking the organizational value.

Right now the first thing to do with AI is to prompt, read, prompt again. It’s how we started with the Skills… it’s a straight line that stops when you stop typing.

What we want, and what Glare Skills should do, is keep running: small stages chained in sequence, each one a place you can inspect and tune. Radar indexes and labels links in one stage, then a separate daily step groups them into topics. That is a sequence, not a single call, and it is the same shape I have been trying to give our Skills so they hand off cleanly from one to the next.

Where we sit in the loop

The other lesson is where the human sits (or little old me). We found two points worth keeping a person:

  1. choosing which topics deserve coverage
  2. judging whether the writing clears the bar.

Everything else runs automatically. That maps onto how I want the Skills sequenced. Automate the mechanics, but leave deliberate openings where judgment enters the loop and feeds the next stage.

@ben and I walk through the project and some of the details in the editing:

The ranking is the clearest example. Every topic starts at 1500, then an arena asks which of four you would read. Your pick moves the scores, and that signal feeds back in. Human judgment becomes a signal the system captures and reuses, which is exactly what I want our Define Skills doing: turning a call into a signal the next Skill can act on.

Where our skills are going

If we get the sequencing right, the Skills stop being one-off prompts and start behaving like a living system. In this way, the mechanics get handled by the robot and we’ll be able to focus on the product and design decisions that are most important.

Radar is live at http://radar.zurb.com. Happy to get into the sequencing conversation with people, we’re in the deepend right now.

1 Like
2

It’s interesting in the sense that you have to still compose problems in a way that you’re not just solving, you’re learning and maintaining what’s happening over the course of a project.

The sequencing that Bryan is calling out, highlights the problem of people jumping into straight execution without understanding what needs to happen throughout a process. If you don’t have a scenario, and a desired output in mind, it’s really hard to get to point A to point B if you never know where to begin, and what it takes to get to something of high-quality.

Hard to imagine how one could do this without previous know-how and experience, simply just knowing that we had to iterate a few dozen times to get this to a place where it both serves and functions.

One thing that I wanted to also highlight was how fun the theming/branding of this was!

3

Agreed. I feel like there are two different systems at work.

Learning loops require intent. They are helping us test assumptions, generate new knowledge, and improve decisions.

Feedback loops help teams maintain quality by monitoring what is happening in the system and signaling when something needs attention.

The goal here is to train AI to be responsive to these so the loops get better.

Trial and error- no better way to learn than just start making things!

Agreed, figuring out how to brand with AI is a challenge. Having fun with these images.

image
image1920×1356 683 KB

4

Piggy-backing off of our post here.

On the engineering side, I think these things become really interesting, and more so as I leverage AI to do a lot of the heavy lifting:

  1. How the architecture is one of the determining factor of quality of the AI outputs.
  2. How much expertise to people actually need to build something like this out? Can anyone do this?
  3. Expertise seems like the barrier to the highest possible quality outcome. OR is it really who has the smartest model, and most amount of tokens?

As an engineer, I’m constantly asking myself, “what should I actually be doing now that AI is such a huge part of my build process?”.

  • Should I be offloading all of the work to someone else?
  • What decisions should I be in control of?
  • Who should be making the decisions, and for what?

One thing that I have found to be true, is that there’s a new level of fun that comes from building out these tools. Almost any feature becomes a, “sure, let’s try that”, instead of the, “uh, that doesn’t seem possible, and it’d take way too long”, just solely from the original pain that would come with building out something complex.

5

Speaking of loops, looks like GPT 5.6 is one of those models that support this kind of thinking. Every.to made a nice article about it and how it supports loops:

If you’ve ever thought about your work, and what you’ve done, there have always been loops.

Loops create consistency amidst chaos. I think we have to start thinking about how we’re deploying our agents in this way.

I’m also realizing how useful AI skills are in creating these loops, started developing a few new ones yesterday:

  • Weekly Summarizer skill - read through all of the code updates (commits, PRs) and summarize what happened last week. This is specifically used for our product rundowns.
  • Daily summarizer skill - take all of the code updates, meetings, notion doc updates, and summarize what happened. Then, look at what wasn’t finished, and what calendar items I have for tomorrow, so that I can thoughtfully prep for the next day.
  • SEO parser skill - consume CSVs or ZIP files of SEO issues (like from Ahrefs), force a more thorough deep dive into the data to confirm what are actually problems, verify them, and then implement changes that solve them properly. Do test-driven with areas that we can.

Super excited to see how we can turn more of this into systems that reduce chaos and increase efficiency.