Tools, Resources, Prompts. Three Words That Changed How I See My AI Agents Forever.

Tools, Resources, Prompts. Three Words That Changed How I See My AI Agents Forever.

Published: February 3, 2026 - 14 min read

In Part 1, I explained what MCPs are and why they made me cry. I walked you through the N x M problem, the USB-C analogy, and the basic architecture of hosts, clients, and servers.

But I left out the most important part.

I told you that MCP servers expose capabilities to AI. What I did not tell you is that those capabilities come in three specific forms. Three primitives. Three superpowers.

Tools. Resources. Prompts.

When I finally understood what each of these meant, I had a realization that stopped me in my tracks.

I have been building AI agents for months. Oprah Winfrey coaches my English eloquence. Tiana Picker coaches my French. Igor Jarvis manages my Substack strategy. Each one has specific workflows, knowledge bases, and instructions.

And suddenly I realized: every single one of them could become an MCP.

Not just agents that live in my Claude projects. Portable agents. Agents that could work with any AI platform. Agents I could share with others. Agents that could outlive any single tool or subscription.

That is what this post is about. Understanding the three superpowers that make MCPs actually useful, and seeing how they map to what you might already be building.

The Three Primitives at a Glance

Before we dive deep, let me give you the overview. Every MCP server can expose three types of capabilities:

That is the high-level view. Now let me show you what each one actually means and why it matters.

Superpower #1: Tools (The AI Gets Hands)

Tools are the most exciting primitive. They are what make AI agentic.

What Tools Actually Are

A tool is a function that the AI can invoke to perform an action. When I say "action," I mean something that has an effect in the real world. Not just generating text. Actually doing something.

Real Examples of Tools

Let me show you what tools look like in practice. And I do not have to imagine, because I already built one.

Back in December 2025, I launched the Alex Bennett LinkedIn Copilot, an MCP server with 16 purpose-built tools for LinkedIn content creation. Here are some of the tools I designed:

Content Creation Tools:

• create_carousel - Generates multi-slide carousels with images and PDF output
• create_visual - Creates graphics in different design styles
• create_quote - Designs quote graphics for thought leadership
• format_text - Applies formatting (bold, italic) for social platforms

Analytics Tools:

• track_post - Logs a published post for performance tracking
• update_metrics - Records engagement numbers (impressions, likes, comments)
• get_stats - Retrieves analytics dashboard with insights

Configuration Tools:

• set_content_path - Configures where files get saved
• set_brand_colors - Sets your brand's color palette
• save_draft - Saves work-in-progress content

Why Tools Matter

Here is the key insight: Tools have side effects. They change state.

When my Alex Bennett MCP calls create_carousel, a PDF file actually gets created on your computer. When it calls track_post, data actually gets saved to a database. When it calls set_brand_colors, your configuration actually changes.

This is what makes AI agents actually useful. They can do things, not just talk about doing things.

Think about the difference:

Without Tools: "I recommend you create a carousel with 5 slides covering these points..."

With Tools: "I have created your carousel. Here is the PDF file, ready to upload to LinkedIn."

That is the difference between an AI that advises and an AI that acts.

The Permission Model

You might be wondering: "Isn't it dangerous to let AI take actions?"

This is where the permission model comes in. When an AI wants to use a tool that has side effects (creating files, modifying data, sending messages), it asks for your approval first.

In Claude Desktop, you will see a prompt like: "Claude wants to use the create_carousel tool. Allow?"

You are always in control. The AI proposes. You approve or deny. This is the safety layer that makes tools practical.

Superpower #2: Resources (The AI Gets a Library)

If tools are the AI's hands, resources are the AI's reference library. They provide context without taking action.

What Resources Actually Are

A resource is a read-only data source that the AI can access. The key word is read-only. Resources do not change anything. They just provide information.

Real Examples of Resources

Alex Bennett MCP does not just have tools. It also exposes 6 deep research resources that I curated. Here is what I built:

Research Resources:

• research://algorithm - How the platform's algorithm works, ranking signals, engagement patterns
• research://job-seeking - Strategies for job seekers, recruiter optimization, portfolio content
• research://b2b-sales - Lead generation tactics, outreach messaging, sales content strategy
• research://thought-leadership - Building authority, personal branding, expertise positioning

Reference Resources:

• templates://post-formats - Different post structures (story, listicle, how-to, hot-take)
• config://brand - Your saved brand colors, fonts, preferences

How Resources Work in Practice

When you ask the AI to help you create content, it can access these resources to inform its recommendations.

For example, if you say: "Help me create a LinkedIn post for job seekers"

The AI might:

1. Access research://job-seeking to understand recruiter optimization strategies
2. Access research://algorithm to know what post structures perform best
3. Use that context to craft a post that is both strategically sound and algorithm-friendly

The result: The AI's suggestions are not generic. They are informed by curated, specific knowledge that you (or the MCP creator) have assembled.

Why Resources Matter

Here is the key insight: Resources let you give the AI expertise it does not have.

Claude is incredibly smart, but its training data is general. It knows about LinkedIn strategy, but it does not know the specific tactics that work right now, in February 2026, based on the latest algorithm changes.

Resources fill that gap. You can load curated, up-to-date, domain-specific knowledge into the AI's context. The AI becomes an expert in whatever you give it resources about.

Think about the difference:

Without Resources: The AI gives generic advice based on its training data from months or years ago.

With Resources: The AI gives specific, current advice informed by curated research you have assembled.

This is how you turn a general-purpose AI into a specialized expert.

Superpower #3: Prompts (The User Gets Shortcuts)

Prompts are the most underrated primitive. They are what make MCPs usable by normal humans.

What Prompts Actually Are

A prompt (in the MCP sense) is a pre-written template that guides the AI through a specific workflow. It is like a macro or shortcut that combines tools and resources into a streamlined experience.

Real Examples of Prompts

Alex Bennett MCP includes 10 workflow prompts that I designed to make LinkedIn content creation effortless. Here are some of them:

Content Creation Prompts:

• viral-post - Create scroll-stopping content with research-backed hooks
• carousel-outline - Plan a multi-slide carousel with hook and call-to-action
• weekly-content-plan - Generate 5 posts (Monday through Friday) with strategic variety

Optimization Prompts:

• analyze-performance - Review your recent posts and identify patterns
• repurpose-content - Take existing content and adapt it for different formats
• engagement-response - Craft thoughtful replies to comments

How Prompts Work in Practice

When you select a prompt, it triggers a pre-defined workflow. The prompt tells the AI exactly what to do, in what order, using which tools and resources.

For example, the viral-post prompt might work like this:

1. User selects the prompt and provides a topic
2. The AI automatically accesses the research://algorithm resource
3. The AI crafts a post following the research-backed structure
4. The AI uses the format_text tool to apply proper formatting
5. The AI presents the result ready for review

The user does not need to know about tools or resources. They just select "viral-post," provide a topic, and get a polished result. That is it.

Why Prompts Matter

Here is the key insight: Prompts turn complex capabilities into one-click workflows.

Most people will never think in terms of "let me access this resource and then invoke this tool." That is developer thinking. Normal humans think: "I want to create a LinkedIn post."

Prompts bridge that gap. They package the technical capabilities into user-friendly shortcuts.

Think about the difference:

Without Prompts: "Use the get_research tool to access algorithm data, then analyze the patterns, then use create_text_post with the story template, then apply format_text..."

With Prompts: Select "viral-post" from a menu. Done.

This is how MCPs become usable by everyone, not just developers.

How the Three Primitives Work Together

The real power emerges when you combine all three primitives. They form a complete system:

┌─────────────────────────────────────────────────────────┐
│                     USER                                │
│                       │                                 │
│                 selects PROMPT                          │
│                       │                                 │
│                       ▼                                 │
│  ┌─────────────────────────────────────────────────┐   │
│  │                   PROMPT                         │   │
│  │         (orchestrates the workflow)              │   │
│  │                       │                          │   │
│  │         ┌─────────────┼─────────────┐            │   │
│  │         ▼             ▼             ▼            │   │
│  │    RESOURCE       RESOURCE        TOOL          │   │
│  │   (provides       (provides     (takes          │   │
│  │    context)        context)      action)        │   │
│  │                                                  │   │
│  └─────────────────────────────────────────────────┘   │
│                       │                                 │
│                       ▼                                 │
│                    RESULT                               │
└─────────────────────────────────────────────────────────┘

Here is a concrete example:

1. User selects the weekly-content-plan prompt
2. Prompt tells the AI to:
   • Access research://algorithm (Resource) for platform best practices
   • Access research://thought-leadership (Resource) for authority-building tactics
   • Generate 5 post ideas
   • Use save_draft (Tool) to save each one
3. Result: User has 5 draft posts saved and ready for the week

The user did one thing (selected a prompt). Behind the scenes, resources provided context and tools took action. That is the power of the three primitives working together.

The Realization That Changed Everything

Now let me tell you why learning this changed how I see my AI agents.

I have been building agents for months. Each one has:

• Specific actions they perform (Oprah evaluates my speaking, Tiana corrects my French, Igor extracts Substack notes)
• Knowledge they consult (evaluation criteria, grammar rules, content patterns)
• Workflows they follow (two-pass review, scheduled posting, pattern analysis)

When I learned about Tools, Resources, and Prompts, I realized something:

I have been building MCP-like systems without knowing it.

My agents already have "tools" (the actions they take). They already have "resources" (the reference material they use). They already have "prompts" (the workflows they follow).

The difference is that my agents are trapped inside Claude Projects. They only work with Claude. They cannot be shared. They cannot be used with other AI platforms. They are not portable.

MCPs could change that.

What if I converted Oprah Winfrey into an MCP? Her evaluation criteria become Resources. Her feedback workflow becomes Prompts. Her ability to analyze transcripts becomes Tools. Suddenly, she works with Claude, ChatGPT, Gemini, or any MCP-compatible client.

What if I converted Igor Jarvis into an MCP? His content extraction logic becomes Tools. His knowledge of my blog archive becomes Resources. His scheduling workflow becomes Prompts. Suddenly, anyone could use Igor for their own Substack strategy.

That is the potential I see.

Not just personal AI assistants. Portable AI assistants. Assistants that are not locked to any one platform. Assistants that can be shared, sold, or gifted. Assistants that survive beyond any single subscription or tool.

Why Most Tutorials Get This Wrong

I want to call something out.

Most MCP tutorials focus almost entirely on Tools. They show you how to build a server that exposes functions the AI can call. And then they stop. That is it.

That is only one-third of the story.

Resources are equally powerful. They let you inject domain expertise into any AI model. They turn generic assistants into specialized experts.

Prompts are what make MCPs usable. Without them, users have to understand the technical primitives. With them, users just select from a menu.

If you only build Tools, you have built a collection of functions.

If you build Tools + Resources + Prompts, you have built a complete experience.

That is what separates useful MCPs from forgettable ones.

The Table Summary

Let me give you the complete comparison:

Your Homework (Yes, Really)

Before Part 3, I want you to do one thing.

Think about something you do repeatedly with AI. Maybe it is:

• Writing emails in a certain style
• Researching topics in a specific domain
• Creating content following a particular format
• Analyzing data with consistent criteria

Now break it down:

• What actions would the AI need to take? (These are your Tools)
• What knowledge would the AI need to consult? (These are your Resources)
• What workflow would tie it all together? (This is your Prompt)

You do not need to build anything. Just think about it.

Because in Part 3, I am going to show you something bigger. Something I predicted back in January that suddenly makes a lot more sense now.

What Is Coming Next

Now you understand the three superpowers: Tools, Resources, and Prompts. You see how they work individually and how they combine into complete experiences. You might even be imagining how to apply them to your own workflows.

But I have not shown you the biggest implication yet.

In January, I made a prediction. I said that AI would become a human-to-human bridge. I said that experts would package their knowledge into MCPs and sell access to their wisdom 24/7. I said that influence would be measured not by followers, but by how many people create agents inspired by you.

In Part 3, I am going to connect the three primitives to that prediction. I am going to show you exactly how Tools, Resources, and Prompts make the human-to-human bridge possible. How a fitness coach's methodology becomes Resources. How their workout generation becomes Tools. How their client intake becomes Prompts.

The vision meets the implementation.

And in Part 4, I am going to take you behind the scenes of the Alex Bennett MCP I have been showing you throughout this post. A complete case study of how I designed all 16 tools, 6 resources, and 10 prompts. The full architecture. The decisions I made. The lessons I learned.

These superpowers are why I predicted something big is coming. Now I understand why.

A Note on This Series

This is Part 2 of a 4-part series called MCPs for Humans.

• Part 1: What MCPs are and why they matter
• Part 2 (this post): The three superpowers (Tools, Resources, Prompts)
• Part 3: How MCPs make my Human-to-Human Bridge prediction possible
• Part 4: Behind the scenes of Alex Bennett MCP (deep case study)

If you have not read Part 1, I recommend starting there. It covers the foundational concepts that make this post make sense.

As always, thanks for reading!