improve

[00:00] So, Shadcn, not too long ago, put out this skill called Improve. If you’re not familiar with Shadcn, they are the UI people. They have a whole component library based on Radix components. Either way, they have this skill. I thought it was really interesting, so we’re going to take a look at it. The skill is called Improve.

[00:25] And what they say is, I guess the general idea is, yeah, it’s right there. The idea, use your most capable model to generate a whole plan, like audit your code. And then you can use a cheaper model to execute those plans. And what appeals to me about this is it’s kind of the way I work with AI. I try to generate like really, really specific things, input schemas and output schemas for any AI I’m working with. So, I just like this as an idea.

[01:03] So, if we go through, let me jump into the actual skill itself. So, they have a bunch of rules, and then they have this workflow. So, there’s a recon phase. There’s an audit phase. And there’s a prioritization phase. And then they write the actual plans out. And the idea is that you should be able to take this plan and give it to some other agent that maybe knows nothing about the project. They do say, where is it? Write the plan for the weakest plausible executor.

[01:32] Now, reading through this, I don’t think they mean, you know, I’m going to throw a little 4 billion parameter model at this thing. So, what I was going to do is just burn up any free tokens I have with various services. So, have one write the plan and then jump over to OpenCode or Antigravity or wherever I have some free usage. And just see how it goes. So, we can install the skill with npx skills. add, shadcn/improve. I’ve already got it installed, so I’m going to skip that part. I’m going to jump over to this very messy project I have called Zaphod. I’m going to open that up in my editor, which I have here. We should end up with a directory of plans here.

[02:19] So, what I’m going to do is I’m actually going to try something called Mimo. And Mimo is, I’m only using because it’s free right now. It’s one of the many, many services that offers like a free, limited free time thing. You don’t have to sign up for it or anything, I don’t think. So, if you want to try something cool, try this guy out. So, this is going to be like my big model. Like, I’m going to throw this at it. It can take as long as it wants or until mimo decides to cut me off or whatever. I don’t really care. So, I’m just going to run skill, improve. Now, I do have an agent’s file in here. And it’s going to, it is reading my agent’s file. And my agent’s file is super simple. It basically just says you have no memory. And you need to get everything from these three files in your memory bank. And this is just how I kind of do it. So, cool. We are going to let this guy run for a while.

[03:25] So, this guy has been running for, I should have kept better track, but we’re approaching 10 minutes. So, I’m going to say between seven and nine minutes. And it’s been parsing through a whole bunch of stuff. And it’s finally coming back with this plan and, or I’m sorry, the results of all the things that it found, all the different little issues and areas. And it is suggesting that we prioritize only creating the plans for a handful of them. I’m going to go ahead and say let’s just create them all. And we’ll let that guy run and we’ll take a look at the plans.

[04:01] All right. So, we’ve got this little plans directory. And if I jump in here, I can see numbered plans from 001 to 010. If we jump into the first plan, it’s established this test framework. It’s got executor instructions, some drift check to see the status of the files it’s going to be working on. And this is really long. Like, I don’t know that I would throw this at a small local model at all. If I jump down to the bottom of this, in plans, there’s a readme. And this explains all the different priorities and the relationship between the two different, or I’m sorry, between all the different tasks. So, 002 depends on 001. Same with 003 and 4 and 5 and 6. They all depend on 1. 10 depends on 8. 9 depends on 1 and 8. Anyway, so there’s a bunch of stuff in here. And we are going to jump over to another service that shouldn’t have any information. We’re just going to tell it to implement this.

[05:17] So, we’re going to jump into OpenCode. And I’m going to switch up my model. We’ll just try Big Pickle. And I’m going to say Implement Plans 001. And we’re just going to run that and see what happens. So, I got to say, so far, I’m really impressed with how this is going. I have not seen the end results yet, but I really enjoy seeing it do this step-by-step process. That is, like, near and dear to my heart. And even though I thought the prompt was really, really big, and I probably would have broken it down into several smaller steps, this particular model seems to be doing really well with this format. So, so far, so good, as far as I can tell.

[06:21] So, it finished up. It took 12 minutes. It got really hung up on testing the database. But what I’m going to do is clear this out. I’m on its branch. I’m going to check the status there. So, it did leave these two unstaged commits. This is actually unrelated to the improve thing. So, I’m going to say this is okay and commit it for it. Okay. So, those are in there. And we’re going to jump over here and look. So, in right here, it marked this task as done. And if I jump back over here and say, bun run test, all of our tests are passing. I’m going to take a minute and review some of this code. And then I think we’ll try another one.

[07:17] So, that actually went really well. I’ve merged that PR or that branch into my main. And I’m probably pressing my luck here. But I noticed that task 4 is really small. It’s just a little bug fix. So, I’m going to press my luck. And we’re going to try this with a local model. Let’s see. I’m going to switch this model. So, I’m in PI and we’re going to switch to this 4-bit quantization. I’m going to say, implement 04. And we’re just going to see what happens. Like I said, I think I’m pressing my luck here. But we’re going to give it a shot.

[08:05] So, it’s wrapping up here. Doing some house cleaning. It didn’t create a new branch, which I’m not super happy with. But I am really impressed with how well the AI was able to follow these instructions. So, at the moment, I’m going to say this improved skill is really good. I think it’s really cool. I think if you want something to audit your code base and give you some direction on things that need to be fixed and present or produce those directions in prompts that you could pass on to another AI. I do feel they’re a little big. But other than that, I think this improved skill is something I will be using going forward. It’s really, really handy. Even if you just want to get, like I said, like an audit of your code base that’s very robust. I mean, it really covers a lot of ground. So, yeah. Check out the improved skill from Shadcn.