PERL: The Smart AI Workflow

**Seamus** (0:11)
All right, welcome back. Another episode of season three with the SEEK Bytes engineering podcast, for engineers by engineers. Today, we're gonna be talking a little bit about something I like to call Pearl. You might know it by different acronyms. We'll get into it in a little bit. But I'm joined here today with Elliott and Will.

**Will** (0:27)
Hello. Hello.

**Seamus** (0:28)
Today's topic is what I'm calling Pearl. This is not my own idea. This is just a easier acronym that I've come up with to help memorize this, because the readings that I've done on this, I can never remember the acronyms. But essentially what Pearl is, is plan, execute, review and learn. And what this is, is a kind of a methodology for using AI agents and getting the benefits out of them while also feeding back the learnings that you get from them back into what we call back pressure. We'll talk a little bit about back pressure in a moment. But I'm curious what you two have been doing with your agents. How have you two been using agents? What's your kind of methodology? Do you have a methodology? Because what I've seen, at least from like the mid junior engineers that I've been working with, a lot of them are just doing one shots where they're in agent mode. They may be not even thinking about the agent that they use, like the model that they're using. They've just got the model set to default. They've got their ID set to agent mode, and they're just throwing in a single prompt and hoping for the best.

**Elliott** (1:24)
Yeah, really great question. So in terms of how I'm using it, there's a couple of different use cases. So I'm getting it to do some planning and kind of analysis because I'm like, I want you to come up with information about something. So an example I used the other day was I wanted to replace an API endpoint with a new one. And so I said, I want you to come up with a plan based on the data you can see flowing through the system and the existing types. I want you to come up with a plan for a new endpoint to handle the same use case, but removing all redundant fields. And so it kind of went through and I was like, great. So it came up with a really good plan. That was really, really useful.
Another use case was actually off the back of something that Seamus did. So we do a quarterly kind of tech health analysis of our different systems to like track speed and agility, the supportability and whatnot. And some of that takes into account purely what's in the code base, different RFCs and whatnot, but also external factors that aren't visible to code such as like, do we have enough staff to support this who know the kind of system? Is it documented well? Stuff like that. And Seamus put together a skill that I was able to use on a bunch of different systems to try and get information which was really useful again. It was supplementary to what I was doing, so I didn't just kind of take it and dump it in there. I'd already gone and done one, but I was like I wanted supplementary information to support the argument that I was kind of making for various levels.

**Seamus** (2:54)
Yeah, nice. I did actually follow Pearl when I was doing that skill. It was kind of my proof of concept. It was kind of me forcing myself to follow this process and actually getting concrete learnings for myself around how valuable it is to follow a process and not just do one-shot things. But before we go into talking about that skill and how I used Pearl to build that skill and improve it, I'm curious with the planning that you did with that agent before, was that just switch to plan mode and just one-shot? Like you gave it a single prompt or were you going back and forth with it on making the plan?

**Elliott** (3:25)
I actually did two. So I did one that was probably more of the P and E of Pearl, where I just said, do this, and I hadn't changed it to plan mode. So it did the plan, and then it immediately started executing. It was following a pretty reasonable plan, but I was like, I don't actually need you to execute because we don't need it now. But I was like, I want the plan. So I did that and I viewed what it came out was. And I kind of like stored that away in working memory. I was like, okay. And then I was like, so delete all that, undo everything you've just done, reset to master.