What AI Agent Skills Are and How They Work

[00:00] What are AI agent skills and 

[00:04] standard adopted by practically 

[00:09] Well, because skills address a 

[00:14] Now AI agents, they're pretty good reasoners and  

[00:19] LLMs or large language models 

[00:24] They can tell you about Kubernetes 

[00:30] or the airspeed velocity of an unladen swallow, 

[00:36] They lack procedural knowledge.

[00:41] The stuff that's specific to 

[00:47] like let's say a 47-step workflow for 

[00:57] Yeah, that would be fun.

[00:59] An AI agent that is running a 

[01:06] encounters a task like generating this 

[01:12] Either somebody needs to prompt it with every 

[01:20] to do that every time, or worse still, the 

[01:26] Now a skill is how you actually 

[01:34] into the agent and the format of a 

[01:41] It's simply a skill.md file.

[01:46] That's a markdown file in a folder.

[01:50] So let me draw out what a 

[01:55] So at the top.

[01:58] Skill.md file is some YAML front letter.

[02:04] So let's have a look at what 

[02:09] Well, at a minimum, there are two things.

[02:13] So there is a name and there is description.

[02:21] These are the two mandatory fields.

[02:24] Now the name identifies the skill,  

[02:28] the description that tells the agent what 

[02:35] Now this description is pretty important 

[02:42] tells agent exactly when this skill applies, 

[02:50] And the description here that says something like, 

[02:57] Now there are some other fields you can put 

[03:02] but it's name and description that are mandatory.

[03:05] Now below the front mater, we also have a field.

[03:12] Now these are the actual instructions.

[03:16] These are the step-by-step workflows, the 

[03:20] whatever the agent needs to know to do the job.

[03:23] And it's just written in plain markdown.

[03:27] And then the skill folder can contain 

[03:32] So you don't have to have 

[03:36] One of those optional folders is the scripts.

[03:42] And this has executable JavaScript or Python 

[03:51] There's also a references directory that 

[03:58] gets loaded if the agent determines it needs it.

[04:02] And finally, the other optional 

[04:07] that contains static resources 

[04:12] That's what an agent skill looks like, but 

[04:20] So what happens when there are 

[04:25] Loading all of them into the LLM 

[04:29] blow through the token budget before 

[04:34] So skills use something 

[04:41] And progressive disclosure works in three tiers.

[04:46] So tier one is metadata only, and that applies.

[04:53] So at startup, the agent loads just the 

[04:58] So that's just a handful of tokens per skill.

[05:01] So even if there are a hundred skills installed, 

[05:07] And this is kind of essentially 

[05:12] Now tier two, this relates 

[05:17] When the agent sees a request that 

[05:21] it reads the complete skill.md body into context.

[05:26] And this tells the agent what to 

[05:31] And that identification, the matching 

[05:36] the skill available is something that 

[05:41] The model decides when it 

[05:45] which is why a good skill 

[05:50] Then tier three, that's 

[05:54] So these are the resources that maps to the 

[06:00] And they only get loaded when a 

[06:04] So the agent starts with a lightweight 

[06:09] That's the name and description.

[06:11] It pulls in the detailed instructions 

[06:15] based on matching the trigger condition.

[06:18] And it grabs resources only at the point of need.

[06:21] Now skills bring a type of knowledge to agents.

[06:26] There are several ways to 

[06:28] So let's just briefly compare them 

[06:34] And the first one I just want to mention is MCP.

[06:39] That's Model Context Protocol.

[06:42] What sort of knowledge does MCP give you?

[06:45] Gives you tool access.

[06:47] It gives agents the ability to call out to 

[06:53] MCP is about what the agent can reach,  

[06:57] but it doesn't tell the agent when to 

[07:03] So that's MCP.

[07:04] Another one is RAG, Retrieval 

[07:10] and RAG that handles factual 

[07:15] Relevant chunks from our 

[07:18] which is pretty handy when the 

[07:22] But RAG doesn't teach an 

[07:24] It's reference material.

[07:27] What about another one?

[07:28] How about fine tuning?

[07:32] What can that do for us?

[07:33] Well, fine tuning bakes knowledge 

[07:40] Now that's something that's 

[07:43] And if the model changes the 

[07:46] Now skills don't really do any of this so 

[07:58] Well skills handle as I mentioned 

[08:05] It's how to do things in what 

[08:10] and because they just files Well, 

[08:14] they could be easily updated and you 

[08:19] Now, in practice, skills will 

[08:24] forms of knowledge like, well, MCP for example.

[08:27] So MCP provides the capability to invoke something  

[08:31] externally and the skill provides the 

[08:37] Now, one more thing to say about skills.

[08:41] Is that the skill.md format is an open standard,

[08:45] and it's published at agent skills.io 

[08:56] and it was adopted across a bunch of AI platforms  

[09:00] like Claude Code and OpenAI 

[09:05] So a skill built for one platform works 

[09:11] Now there's a useful way to think about 

[09:16] Now humans have distinct types of memory.

[09:19] There's semantic memory, which are facts.

[09:22] So Rome is the capital of Italy.

[09:25] There's episodic memory, which are experiences.

[09:29] So, uh, I went to Rome last summer.

[09:33] Actually I did, and it was lovely.

[09:35] Uh, and then there's procedural memory, which 

[09:41] streets of Rome and live to tell 

[09:47] Now agent architectures are 

[09:52] So semantic memory, that maps 

[09:59] augmented generation and knowledge bases.

[10:02] Episodic memory.

[10:03] Well, that really maps to conversational.

[10:08] Logs and interaction history 

[10:12] Well, yep, that maps quite nicely to skill files.

[10:20] Now, one thing that does need mentioning 

[10:29] scripts with access to file systems 

[10:37] That's what makes them powerful, but 

[10:45] Because when an agent runs one of these scripts,  

[10:50] it's typically executing 

[10:53] and audits have found publicly 

[10:58] contain bad stuff like prompt injection, bad stuff

[11:04] like tool poisoning, bad 

[11:10] Basically the usual suspects 

[11:13] So, so treat skill installation 

[11:18] team treats installing any software dependency,

[11:21] which is to say, review it and understand what 

[11:28] So, So where does this leave us?

[11:30] Well, skills are procedural memory for AI agents.

[11:34] They're defined in a markdown 

[11:38] that teaches an agent how to do a specific job.

[11:42] Skills are conditionally triggered 

[11:48] progressive disclosure and the 

[11:52] So an agent that already 

[11:56] of an unlaid and swallow, African and European,

[11:59] can now also learn how to perform any 

[12:05] So that's AI agent skills.

[12:07] If you're using them, let me know in the comments.