Multi SKILL.MD Configurations: Self-learning AI — Transcript

Hello community, so great that you are back. Let's talk about multi-skill markdown configurations.

Yes, I have a video here for single skill MD, but let's have a look how we combine it.

What I really enjoyed was this publication here from February 2026 here, where agentic skill for LLMs, where the authors showed that rather than encoding all procedural knowledge within a model weights, the agent skills, this means composable packages of instructions, code, and resources that agents load on demand, enable dynamic capability extension without any retraining of our AI models.

Now, you are familiar here with our level one metadata, level two instruction loader or trigger, and level three resources with dynamic loading, scripts, references, asset templates, we have seen this, and you load everything into context window.

Great. Now, just to be sure where we are.

In my last video, I talked to you about in-context reinforcement learning in an agentic way, so this is not what we are talking today.

This was yesterday, this was a complete different complexity.

Today, we are just talking about this, we are talking about skill MD, this means in-context learning.

Great. So, there's no generalization of those skills of your AI.

You train on a particular skill, it is an in-context learning, and you cannot assume that your AI, given whatever intelligent it has, is not able to have generalization or out of domains solver.

So, you have a context window, let's say you have 200,000 tokens, you upload a skill A from some marketplaces, a skill B, and then you see, ah, for the skill A, there's a much better solution, or there's another methodology, so if you load all these skills and you have no idea what is the mathematical solution or the solver behind this, you might want then to eliminate some skill that are all within the context window, but unlearning in your context window is quite a challenge.

So, let's see if we can find a better solution to this, yeah.

So, therefore, I decided to modify my view of the world, I say in-context learning, beautiful.

But you know, we always should provide some few shot example to our in-context learning, yeah, to show the AI, hey, this is exactly what I want.

Now, we just extend this now and we add here in the same category, skill markdown.

Those description, as I showed you here in my first video on single skill MD, this is exactly more or less here a continuation now on a higher complexity level.

And of course, you know, we have here a sequence of actions, a sequence of MCP, or if you go classical, API calls, whatever there's necessary to solve it.

But you remember, I showed you just in my video here, hopefully this was yesterday, when the AI agents get locked in.

They cannot solve it anymore, there are some configuration where the agents just stop to perform the job.

So, can we solve this in addition?

And as I told you, skills and MCP are not competing standards, but more or less think about orthogonal layers of the emergent agentic stack, and here you have the agentic skills and here MCP, beautiful.

So, a skill might instruct the agent to use a particular MCP server, specifying how to interpret its outputs, and define fallback strategy if the connection fails.

Skills provide the procedural intelligence, MCP provides the connectivity.

And I think they go very nicely together.

Now, you have Claude Code, you have skill creator meta-skill where you just scaffold your new skill from a natural language description, generating the directory structure, skill.md, and bundled scripts.

And everything is hopefully working out if you have a real simple problem.

But we also have enterprise deployments at companies such as Atlassian, Canva, and Sentry have produced production-grade skills that encode proprietary workflows.

Great.

Now, you know from my single skill MD, our skill MD, our reference MD, our forms MD, everything beautifully.

And you remember that we had to look here at the Anthropic financial service plugins here for a competitive financial analysis, the skill MD file.

It had beautiful instruction, if the prompt lists seven competitors, include all seven, not five or six.

And if the prompt shows data for year 2015 to 2026, include all years, not a subset.

So, you immediately got a feeling, hey, this is going to be fun if we build now multi-skills.

So, what we are really facing is a problem here that humans write skill, this is nice, no, this is beautiful, okay.

And then we have a little bit of AI machines like Claude Code or so, they write skill, and this is really nice, but if you have here the access of the complexity, and let's say you have to write a complex industrial solution, or you're working in science and you're working in nanotechnology, or in biotech, or for the next chip design, you don't have single simple skills.

You have, my goodness, a complexity that we are going to talk about in this video.

So, welcome, let's start.