In the movie Matrix, there is this great scene where Neo points to a helicopter and asks Trinity "Can you fly that thing?". Trinity responds "Not yet", and asks the operator to remotely upload the skill to her brain. Moments later she goes, "I can fly that thing now".
The new "agent skills" protocol for LLMs reminds me of that scene.
What Are Agent Skills?
The idea is simple: domain-expertise can be loaded on-demand by the LLM at run-time, directing itself without the user having to provide the context.
A Real Example: Denver Weather Analysis
Here's a practical example from my recent work. We have been having pretty warm weather in Denver. I find myself commenting "this is highly unusual". But I wanted to know if that's just a feeling or backed up by data.
I ran a first analysis and shared it. I got suggestions from a few knowledgeable people. There were important missing pieces that I didn't know:
- "Sen's slope" (a statistical method for trend analysis)
- The right way to cut the data by seasons
- Focusing on deviations from historical means
- Using joy plots for visualization
Building the Skills Framework
I built an easy-to-use "skills" framework into VerbaGPT, and added the suggestions I received as-is. The results surprised me. Even though I'm not a weather expert, I can still tell that the quality of output was dramatically better.
Why This Matters
I'm naturally skeptical of AI buzzwords, however the agent skills protocol seems highly useful thus far. This idea can be especially powerful when the right context is invoked at the right time in a multi-stage analysis.
The real power of this approach is:
1. Domain expertise becomes reusable - You capture what domain experts know
2. Non-experts can produce expert-level work - The LLM applies the expertise systematically
3. Context is injected at the right moment - Skills activate when needed in the analysis pipeline
The Future
I'll keep testing and sharing what I'm learning here with other interesting data (if you have suggestions, let me know).
And while these charts aren't exactly merry, I do wish you happy holidays!
Full Analysis & Data: Interactive analysis available on VerbaGPT
Data Source: ERA5 monthly averaged data on single levels (2m temperature) from 1940 to present (the data window available from ERA5)
Originally posted on LinkedIn in December 2025. Tagged #CoffeeAndCode