00:00:00This is Skills By The Cell, a directory of agent skills that you can add to any agent
00:00:04to extend its capabilities, like making it amazing at React, turning it into a front-end designer,
00:00:10or even making it a motion graphics editor. The whole thing works by cloning and scanning a repo
00:00:15to see if it has a skills.md file before installing the skill. But will Cloudflare's new proposal
00:00:22make this amazing resource obsolete? Hit subscribe and let's get into it.
00:00:27Agent skills were introduced by Anthropic last year to improve Cloud's capabilities on specific
00:00:32tasks, and works by adding a skills directory to the .cloud one and then adding another
00:00:37directory with the name of the skill, followed by a skills.md file.
00:00:42Now this file needs to contain a few key things for the skill to work. First, the top section,
00:00:47which is front matter in YAML, that should contain at least the name of the skill and the description.
00:00:52Then, below that, outside of the front matter, this is where the instructions for the skill should go.
00:00:58The instructions get loaded into the model's context so that it can be used for the specific task.
00:01:03And the beauty of skills lies in a technique called progressive disclosure, where only the essential
00:01:08information is shown at first, and the model can dig in to find even more if necessary.
00:01:13So when an agent loads up, only the information in the front matter of each skill is loaded into
00:01:19the context, and after it reads a prompt, it checks the description of each skill to find out if a
00:01:24skill can be used to enhance the prompt, and if it finds one, it then adds everything outside of the
00:01:31front matter into the model's active context. You can even link other files to skills,
00:01:36like references, for a model to grab those if it thinks it's necessary.
00:01:40Since then, other companies like Microsoft, OpenAI, and OpenCode have adopted agent skills.
00:01:46But there hasn't really been an easy way to install a skill for multiple agents without manually going
00:01:52into the relevant directory and cloning the repo. This became especially apparent when Vassal wanted
00:01:58to share skills for their tools for people to put in their specific agent. So they made a project
00:02:03called Add Skill, which makes it easy to add a skill to 16 different agents. Simply NPX Add Skill,
00:02:11followed by the name of the skill or the location of the repo.
00:02:14Now I know what you're thinking, what's stopping people from using Add Skill to add any repo?
00:02:19Well, the way it works is after it passes the command, it then clones the repo, then looks
00:02:25for a skill.md file. If it can't find it, it cleans up by removing the repo. But if it can find it,
00:02:32it auto-detects a user's installed agent by looking through the configs, then creates a
00:02:37sim link from .agents/skills to wherever the skills need to be located for that agent. So .cursor for
00:02:44cursor and .clause for Claude code. It also has some telemetry for tracking, which I'll talk about
00:02:49later. But at this point, Vassal had a way for people to easily install skills to any agent.
00:02:55But what about discovering a skill? This is where skills.sh comes in, showing currently a long list
00:03:01of uncategorised and unpaginated skills. And the list gets populated, I think, by someone first
00:03:08installing a skill using Add Skills or Skills, which kicks off the telemetry that's inside the package
00:03:14to add things like name, agents and other information. The telemetry is completely anonymous
00:03:20and is not only used to populate the list, but also the amount of times a skill has been downloaded
00:03:26and by the specific agent, which can also be used to track the trending skills in the last 24 hours.
00:03:32And also, if you're worried about being tracked, you can totally disable the telemetry.
00:03:37It's amazing how much work Vassal Labs have put into Skill Discovery and it's being used
00:03:42by loads of people to discover awesome skills like React Native skills for your agents,
00:03:48best practices for better auth and even the super popular ReMotion skill which people are
00:03:53using to create all kinds of crazy videos just from a single prompt. But has Cloudflare's new proposal
00:03:59made this all for nothing? Because it introduces an alternative way for agents to discover skills
00:04:05by first fetching a lightweight JSON file from a well-known location that lists out all the available
00:04:12skills a place has to offer. So you could essentially write a prompt saying "Build me a Cloudflare project
00:04:18using Wrangler" and the agent will first read the prompt then check the well-known location
00:04:24like cloudflare.com/well-known. Of course it won't actually be well-known it'll be something else
00:04:30/skills and then retrieve the index.json from that location which includes mandatory information per
00:04:38skill like name description and number of files related to that skill. After that it passes this
00:04:44information to the agent and then based on the prompt the agent will realise that the Wrangler
00:04:49skill is needed from this well-known location based on the index.json file. So it will make a call
00:04:55to that location to get all the files related to that skill and cache them in case they're used
00:05:01again in a follow-up prompt but use that skill for its response that it gives back to the user.
00:05:08So while Cloudflare's proposal does sound really cool it does add a few extra steps to making a
00:05:14skill discoverable like adding that index.json file with the relevant information and also adding it to
00:05:20a well-known URL which I'm not exactly sure what that means but compared to putting a skill on GitHub
00:05:27and having it found by Vercel versus this way I think Cloudflare's approach does harm smaller
00:05:33developers because of all these extra steps but if it does get accepted and I'm not also sure who's
00:05:40reviewing this and who has to accept it then it means goodbye to Vercel's skills.sh project.
00:05:46However I'm sure there is a way they could work together because you could use skill.sh to find
00:05:52skills from indie and lesser known developers and you could use Cloudflare's approach for an
00:05:57agent to discover a skill automatically from big companies like Cloudflare and Vercel. Either way
00:06:03with developers using these skills to create amazing new projects you're going to need some
00:06:07error tracking in case they explode without you knowing and this is where better stack comes in
00:06:12because not only does it ingest logs from your back end it can also keep track of front-end errors
00:06:18using its AI native error tracking and to top everything off better stack can also create
00:06:23beautiful status pages so go ahead and check out better stack today.