Chengchang Yu
Published on

MCP + Skills - When AI Finally Learns Like Humans Do

Authors

Remember when you started your first job?

Your manager didn't just tell you what to do - they showed you how the team works, gave you access to tools, pointed you to documentation, and shared best practices from years of experience.

That's exactly what MCP + Skills does for AI.

And it changes everything.

The Evolution: From Instructions to Knowledge Transfer

Before Skills: Teaching AI to Talk

In the past, all we could pass to AI was "how to talk to it" - prompts.

"You are a data analyst. When I give you data, 
analyze it and output in this format..."

Every conversation, you're teaching it the same thing. Like training a goldfish.

With Skills: Teaching AI to Work

Now with Skills, we can pass:

  • ✅ Knowledge (domain expertise, company standards)
  • ✅ Tools (executable scripts, APIs)
  • ✅ Methodology (step-by-step workflows)
  • ✅ Best Practices (lessons learned from real work)

It's not just remembering instructions - it's learning how to work.

The Real Breakthrough: Systematic Learning

Here's what changed fundamentally:

AI finally has a systematic way to learn.

Not just memorizing facts, but learning:

  • How to use tools
  • How to follow methodologies
  • How to apply best practices
  • How to find and reuse existing solutions

A Real Example: AI Looking for Best Practices

Here's something fascinating that happened when using Claude Code with Skills:

Task: Write a script to fetch Youtube video subtitles.

What Claude did:

  1. First, it used MCP to search GitHub
  2. Found existing libraries others had written
  3. Referenced or directly called those libraries
  4. Delivered the solution

This is AI actively seeking best practices.

It didn't just write code from scratch. It did what a senior engineer does: check if someone already solved this problem, then build on their work.

MCP + Skills: The Power Duo

MCP and SKILLs, Image from: 刘一缘

MCP and SKILLs, Image from: 刘一缘

Let's break down why this combination is so powerful:

MCP (Model Context Protocol): The Connection Layer

What it does: Defines how AI communicates with external tools and services.

Think of it as: Teaching AI how to use the phone.

MCP Server: GitHub
↓
Claude: "Search for Youtube subtitle libraries"
↓
MCP returns: [library_1, library_2, library_3]
↓
Claude: "Read library_1's documentation"

Skills: The Knowledge Layer

What it does: Packages the knowledge, tools, and workflows AI needs for specific tasks.

Think of it as: Teaching AI how to do the job.

youtube-research-skill/
├── SKILL.md              # "When researching products, do this..."
├── scripts/
│   └── fetch_subtitle.py # Tool to get video subtitles
├── references/
│   └── analysis_framework.md # How to analyze content
└── assets/
    └── report_template.md    # Output format

Together: AI That Actually Works

MCP without Skills = A phone that can call anyone, but doesn't know who to call or what to say

Skills without MCP = A training manual with no way to access external resources

MCP + Skills = An employee who knows:

  • What tools exist (MCP)
  • When to use which tool (Skills)
  • How to combine tools to solve problems (Skills + MCP)
  • Where to find best practices (MCP search + Skills guidance)

The Four Transformations

This combination enables four fundamental shifts:

1. From Implicit to Explicit Knowledge

Before: "Just figure it out" (hoping AI guesses correctly)

After: Codified workflows, documented standards, structured processes

Example:

# Product Research Methodology

1. Search for top 10 videos by view count
2. Extract subtitles for each
3. Identify common themes across videos
4. Flag potential sponsored content (keywords: "Collaborate", "Sponsor")
5. Synthesize into comparison table

2. From One-Time to Reusable

Before: Write a long prompt every time you need product research

After: Load "product-research-skill" once, use forever

Impact:

  • First time: 30 minutes to set up
  • Every time after: 30 seconds to invoke

3. From Isolated to Connected

Before: AI works in a vacuum, can't access external knowledge

After: AI actively searches GitHub, reads documentation, finds existing solutions

Example:

  • Need to process video subtitles?
  • MCP searches GitHub for existing libraries
  • Skills knows how to evaluate and use them
  • Result: Better solution in less time

4. From Individual to Collaborative

Before: Each person has their own prompts, no consistency

After: Team shares Skills, everyone gets the same quality output

Real-world scenario:

marketing-team-skills/
├── brand-voice-skill/
├── competitor-analysis-skill/
├── content-calendar-skill/
└── performance-report-skill/

Everyone on the team loads these Skills → Consistent brand voice, standardized analysis, unified reporting.

A Real Use Case: Product Research Agent

Let me show you how this works in practice:

The Problem

You want to buy an air humidifier. There are dozens of review videos on Youtube, but you don't have time to watch them all.

The Traditional Approach

  • Spend 1+ hour watching videos at 2x speed
  • Take notes manually
  • Try to remember which video said what
  • Make a decision based on incomplete memory

The MCP + Skills Approach

Step 1: Create the Skill

youtube-product-research/
├── SKILL.md
│   └── "When user asks for product research:
│        1. Search Youtube for relevant videos
│        2. Extract subtitles using fetch_subtitle.py
│        3. Analyze using analysis_framework.md
│        4. Output using report_template.md"
│
├── scripts/
│   └── fetch_subtitle.py
│       └── Uses MCP to search GitHub for best Skills
│       └── Implements subtitle extraction
│
├── references/
│   └── analysis_framework.md
│       └── How to identify key features
│       └── How to spot sponsored content
│       └── How to compare products objectively
│
└── assets/
    └── report_template.md
        └── Structured output format

Step 2: Use It

You: "I want to buy an air humidifier, budget $50-100, for a 300 sq ft room"

What happens behind the scenes:

  1. Skills activates: Claude loads "youtube-product-research" Skill
  2. MCP searches: Finds top 10 Youtube videos on "air humidifier review"
  3. MCP fetches: Gets subtitles for each video (using GitHub library it found)
  4. Skills analyzes: Applies analysis framework to identify:
    • Key features mentioned across videos
    • Common complaints
    • Price ranges
    • Sponsored vs genuine reviews
  5. Skills synthesizes: Generates structured report with:
    • Top 3 recommendations for your needs
    • Feature comparison table
    • Links to original videos for verification

Time: 10 minutes (vs 1+ hour manually)

Quality: Comprehensive analysis of 10 videos (vs 3-4 you'd watch manually)

Step 3: Reuse It

Next month you want to research standing desks. Same Skill, different product. Just ask.

The Paradigm Shift: AI as a Learning System

Here's the profound insight:

Before: AI was a tool you operated

After: AI is a system that learns and improves

Traditional AI Workflow

You → Write Prompt → AI Responds → Forget Everything

MCP + Skills Workflow

You → Create Skill (once) → AI Learns
                          ↓
                    AI uses MCP to find tools
                          ↓
                    AI applies methodology
                          ↓
                    AI delivers result
                          ↓
                    Skill improves over time

The Engineering Principle: Make Tacit Knowledge Explicit

This is what great engineering teams do:

  1. Document processes (don't rely on tribal knowledge)
  2. Standardize workflows (don't reinvent the wheel)
  3. Build reusable tools (don't repeat yourself)
  4. Share best practices (don't work in silos)

MCP + Skills brings these principles to AI.

It's taking the implicit knowledge in your head - "this is how we do product research" - and making it explicit, structured, and reusable.

Why This Matters for Production Systems

If you're building AI systems for real business use, this combination solves critical problems:

Problem 1: Inconsistent Output

Solution: Standardized Skills ensure consistent quality

Problem 2: Can't Scale

Solution: Reusable Skills mean one expert's knowledge serves the whole team

Problem 3: Can't Improve

Solution: Skills can be versioned, tested, and iteratively improved

Problem 4: Isolated from Reality

Solution: MCP connects AI to real-world tools and data sources

The 50/50 Rule Revisited

Remember: AI Agent's future = 50% model + 50% engineering

MCP + Skills is the engineering half.

  • MCP = Infrastructure layer (how to connect)
  • Skills = Application layer (what to do)
  • Together = Production-ready AI system

Getting Started: Your First Skill

Don't overthink it. Start with something you do repeatedly:

Example: Weekly Report Skill

weekly-report-skill/
├── SKILL.md
│   └── "Generate weekly report every Friday
│        Include: key metrics, highlights, blockers"
│
├── scripts/
│   └── fetch_metrics.py
│       └── Uses MCP to query your analytics system
│
└── assets/
    └── report_template.md

Result: Every Friday, just say "generate weekly report" and it's done.

Conclusion: AI That Finally Gets It

For the first time, we can transfer to AI not just what to say, but how to work:

  • ✅ The knowledge (Skills)
  • ✅ The tools (MCP)
  • ✅ The methodology (Skills)
  • ✅ The best practices (MCP + Skills)

This is how humans learn. Now it's how AI learns.

And when AI can learn systematically - not just memorize prompts - that's when it becomes truly useful in production.

The future isn't about smarter models. It's about models that can learn to work the way we work.

MCP + Skills gets us there.

Related Resources: