NEW Laravel Boost v2.0: How Agent Skills Actually Work?

Hello guys. So the version boost 2.0 is out released and in this video I will try it out. I will show you the difference and the new thing about skills and how exactly it impacts the results. First to demonstrate that in action let's create a totally new Laravel project with specifically Livewire starter kit and we'll see how much of Livewire knowledge is packed into the guidelines or skills for the AI agents. So I choose yes for boost and this is the thing that is new in version two. You may choose to install AI guidelines which were there previously and then agent skills. So let's enable all of them. Also you may choose separate providers which in this case is Laravel 45 that comes with starter kit and then later we will add filament here. But here as you can see it's guideline. So there's additional information whether it's a guideline coming from thirdparty package or skill and package authors may actually create and suggest both. I will choose no for herd and for editors I actually use clot code mostly but just for the demonstration let's also choose codeex and this is probably the most important thing changed. So you have guidelines for selected agents which come into claude MD or agents MD or general guidelines combined in one file but then separately three files are registered as skills provided by the Laravel team themselves. So in this current stage they moved some of the guidelines into skills. In one of his tweet replies, pushpack, one of the authors of boost said that they will be improving guidelines and moving more to skill. But yeah, this is the result for now. And how does it actually look in the codebase? So, first the main guidelines file for the agents, which is claude MD and agents.m MD4 codeex. And they claim it to be shorter. So, how many lines? These are 292 which is impressive because I remember in boost one it was more like 500 to 700 lines of markdown depending on which packages I choose to install. But also what is interesting in the boost guidelines there's a new section skills activation. And here's the list of instructions when to activate the skills with quite a lot of text. And we will actually test whether those skills are activated automatically or we need to activate them manually. Those skills are now separated in doclude for example skills and then those three folders and each folder contains skill.md with instructions. It's basically the same probably the same I assume that was in claude MD but then we have name and description as a skill structure which means that when you prompt AI claude MD will be loaded automatically into context into memory and skill MD not necessarily it will be called only when needed which means saving context space for LLM for better results then for example let's add filament to the Laravel project and see how boost handles it. So we have it installed and now let's run PHP artisan boost install again. So it would detect filament in action. We repeat the choice and here you see filament in the list. Let's add that. And at this point it will override the claw and agents again reasking us if we want to make any changes. And now as you can see there are 14 guidelines including filament here but also one more skill was detected which is tailwind CSS development. Not sure why filament installation triggered this one to install but probably useful especially since skills don't take extra context and they are separated from the main guidelines. But now if we take a look at those guidelines cloud MD should have filament section somewhere here. Yes. So we have filament filament rules that come from the guidelines from filament package creators. And now in total our guidelines is 400 lines of code. So you see with each package with each set of instructions including your own maybe custom instruction from folder the main cloud MD is kind of polluted in a way with useful information. Yes. But for a specific prompt, there would be only like 5% of that information actually usefully used. But regardless of that, all the cloud MD would be loaded into memory. Inefficient, isn't it? And now let's see the skills in action. Will they be automatically called for that? I will launch Codeex, specifically Codeex and not the popular cloud code, just out of curiosity whether they will be called from Codeex as well. I actually have subscriptions to both cloud code and codeex and switching between the models when needed. So we will use GPT52 codeex and we will prompt for fresh simple API and will it call past tests skill and this is my prompt create model migration for post create factory and cedar and create API endpoint to get all posts I don't mention tests specifically but somewhere in agents MD and cloud MD I remember it requires to generate the test along the way with a new feature let's see and it seems to be done within 4 minutes and I see that the tests were executed but was it with the skill let's ask codec so we can prompt after the successful prompt what did you actually do along the way and it admitted there were no skills activated and interestingly since past tests were executed and created I was expecting that to automatically trigger the skill because in cloud MD and in agents.md we have this past testing when writing tests, creating unit or feature tests and so on. But apparently it's not enough in cloud MD and agents MD or maybe it's because of codec. I will retry the same in cloud code in a minute. And I found that in the official skill support pull request by pushback he emphasizes that skill activation should be done something like this. Activate skill name to help with something. Let's try to do that. So I have restarted Codex and updated it just in case and I add this to the prompt activate specific skill and the prompt is relatively the same just for pages model instead of posts and along the way I already see called skill MD from codeex skills past testing. So that was successful. Basically what I want to show it doesn't even really matter the result actually in this case. What matters is if you want to enable and activate the skill, the best way, the guaranteed way is to do that manually in the prompt. Technically, the LLM and cloud code or codeex should catch the skill by description of the skill or in cloud MD, but it doesn't always do that. Let's actually try cloud code. And this will be the prompt also create model migration for another thing like tags. I will not mention the skill and I'm using sonet model. And let's see if the skills will be activated. And let's look at the result. All tests passed successfully. But was the skill called? If we go up to PHP artisan test, let me create tests. And I don't see the skill here. And again, the prompt, did you use any skills in last prompt or in last prompt? I probably should put it in singular. Actually, let's reprompt for the correct result in last prompt. And this is the result. pretty typical for LM. No, I didn't use but I should have because the guidelines actively state important in caps in bold in screaming whatever you call that activate every time you're working with past. So it asks would you like me to be more proactive and if I answer yes probably it will increase the probability of skills being called but by default as you can see it doesn't always call the skill unless you specify it to in the prompt. So the upside of skills is that context space is saved. So it's not in the main guidelines in the main cloud MD. So that's a good news. And we can actually run context and see if we go up what do we have here? How much space is used? As you can see skills are using only like 400 tokens. I'm not actually sure which line is cla MD here. Probably in the system prompt I'm guessing. or maybe in the messages. This is the MCP. This is the plug-in. All right. Memory files 5,000 tokens. So, this one is taken by guidelines. So, basically by moving from guidelines to skills, we decrease this part and only slightly increase this part if the skills are called. But then also, we risk skills not to be called and probably add more work to ourselves to prompt actively. what skills do we need to call? We need to remember them if you have a lot of them and so on. So this is kind of a personal preference of choice between guidelines and skills. Pushback has also released an article a post on Twitter. By the way, for some reason this format became more popular articles full text on Twitter instead of blog posts or separate documentation pages. So yeah, he lists guidelines versus skills. What is the actual difference? So guidelines are more foundational and skills should be extra when needed. Something like that. But also I like the tweet by pushpack himself a few days earlier summarizing it basically in a better format. Skills or guidelines you kind of never know. In my experiments with AI, even with the guidelines, you don't have 100% guarantee that some kind of rule or guideline will be actually used. But with guidelines, at least you're sure it's in the memory. it's loaded automatically and with skills it may be not loaded at all but a good example of skill and I tried it on AI coding daily my separate channel is front end design it's a plug-in from official anthropic which is actually a skill of front end design so this is very good I've tried it in this video and I will link that video in the description below so you would understand better what is maybe the usage of totally separate skill for specific better functionality of something also in that article on Twitter pushpack detailized how you can use community skills like this one so it's almost like a trend now for companies to release their skills like versel I've seen posgress using like their best practices skills but I'm not really sure if they are that useful in laral projects but you may try it out there's a command now boost add skill you also create your custom skills in the similar fashion like you had guidelines in boost one now you may AI skills with the same logic. When you run boost install or boost update, the skill will be automatically detected. And if you are a package creator, the skill for your package may be placed in your repository in this file and then it would be detected again on boost install command. So yeah, great release of boost version two. Skills is kind of the number one feature, but there are more things here in the pull requests. We'll actually see in the future if the current trend of skills will continue as kind of a better way to provide the guidelines for AI agents. Maybe the community will come up with something new. The trends are always changing and actually Laravel team with boost and other tools are trying to keep up very nicely. So in Laravel we have firstparty support for things like AI and also Laravel AI SDKs coming quite soon where we will be able to use AI agents in Laravel code calling them calling like open AI API and stuff like that. So I'll talk about that on this channel as well when it's released. And if you're interested in my AI experiments in general, not necessarily with Laravel, again, I have a separate channel AI coding daily and a website aiccodingaily.com with various tutorials and news and prompts and even community tweets I follow. And at the bottom of that, there's a form to subscribe to free newsletter that I send every Wednesday with the summary of that week in AI coding. And I also publish that as an article. So this is an example my videos and then what happened over that week in the AI coding community. So if you want to get that every Wednesday subscribe the form is below on that page and I'll also link that in the description below. What do you think about new Laraveville boost? Will you use it? Do you use skills or do you prefer guidelines? We can discuss all of that in the comments below. That's it for this time and see you guys in other