Carl Yeh

There are some use cases.

there's some use cases that, I mean, sorry, edge cases that we haven't seen before.

And so it's like, oh, in one instance, like if the property code has these extra three letters, what do we do?

I was like, oh, we do this, this, this, but that gets written up in the retrospective and then automatically populates into learning and skill.

So the next time you don't have to remember that it's already part of the learning and skill that,

It's written over.

So, yes, essentially you are self-improving the agent, especially on the skill side.

So you do not have to reprompt any edge case or any specifics outside of like any major overhaul you need to do with this.

Yeah, exactly.

And it's actually not hard to set up because, like...

In this case, I'm using subagents, but let's just, for example, ditch the subagents and you're just in a folder and you're just like, you have Claude create a skill.

I don't know what that skill would be for that specific use case.

And then when you engage in that task, just build out, literally ask Claude, hey, I want to build a slash command retrospective.

And it does these three things, review, books out learning failures, update the skill in the folder.

And that's it.

And so now you have a slash command, then you have a skill and it automatically updates every single time.

You can make it as simple or as complex, a little bit more complex how I'm doing here, where every single agent skill gets updated if it's appropriate, if the learning affects them.

So the skill, so we have like a, let's say property value.

Yeah, let's say, let's look at this description extractor here.

So you have the skill, which instructs the agent to what it actually does.