When most people think of AI, they picture a chatbot — you type something, it types back. That is useful, but it is limited. The AI is just talking. It cannot actually do anything.

AI agents are different. They do not just respond — they act. They can search the web, read a document, look up a record, send a message. The thing that makes this possible is called a skill.

What Is a Skill?

A skill is a specific task that an AI agent is allowed to request.

Every agent starts with a list of skills — a bit like a job description handed to a new employee on day one. The AI reads this list, figures out which skill is relevant to what you asked, and uses it.

Think of it like a laminated card your assistant carries around:

Can do: Search the database
Can do: Book a meeting room
Can do: Send a draft for your review
Cannot do: Approve payments

The AI works within that list. It cannot go off-script.

What a Skill Looks Like

Behind the scenes, each skill is a short description card with three parts — a name, a description, and the details it needs to do the job:

{
  "name": "search_cases",
  "description": "Search legal cases by keyword and location",
  "needs": {
    "keyword": "what to search for",
    "location": "SG, MY, or HK"
  }
}

The description field is the most important part. It is what the AI reads to decide whether this skill matches your request. A vague description leads to the wrong skill being chosen — or none at all.

How the AI Receives Its Skills

When you send a message to an AI agent, your skills list travels alongside it — the AI reads your message and the skill list together, in the same moment:

{
  "message": "Find me recent landlord dispute cases in Singapore",
  "available_skills": [
    "search_cases",
    "open_case",
    "summarise"
  ]
}

It does not browse a separate menu or search an app store. The full list of available skills is placed right in front of it, alongside your question, every single time.

Three Actors, One Conversation

An AI agent is not one thing — it is three actors working together. Understanding who does what is the key to understanding the whole system.

Step	Who	What happens
1	🧑 You	Send your message to the Agent
2	🤖 Agent	Receive your message and package it with the skill list
3	🤖 Agent	Forward everything to the AI
4	🧠 AI	Read the message + skill list and decide what to do
5	🧠 AI	Request a skill with the right details — or reply directly
6	🤖 Agent	Execute the actual task when the AI requests one
7	🤖 Agent	Return the result to the AI
8	🧠 AI	Write the final response and hand it back to the Agent
9	🤖 Agent	Deliver the final reply back to you
10	🧑 You	Receive the final reply

The important thing to notice: the AI never talks to you directly, and never executes anything directly. The Agent is always in the middle, coordinating between both sides.

Here is how a single request flows between all three:

sequenceDiagram
    actor User as You
    participant Agent as Agent (System)
    participant AI as AI (LLM)

    User->>Agent: 1. Sends a message
    Agent->>AI: 2. Packages message + skill list
    AI->>AI: 3. Reads everything and decides

    alt Needs a skill
        AI->>Agent: 4. Requests a skill with details
        Agent->>Agent: 5. Executes the actual task
        Agent->>AI: 6. Returns the result
        AI->>Agent: 7. Writes final answer
    else No skill needed
        AI->>Agent: 4. Writes direct answer
    end

    Agent->>User: 8. Delivers the reply

What the AI Is Actually Reading

When your message arrives, the AI sees something like this all at once:

Available skills:
  search_cases  — find legal cases by keyword and location
  open_case     — read the full content of a specific case
  summarise     — condense a long document into key points

Your message:
  "Find me recent landlord dispute cases in Singapore"

The AI reads this, decides search_cases is the right fit, and responds with a clear, structured request:

{
  "skill": "search_cases",
  "keyword": "landlord disputes",
  "location": "SG"
}

Your system then runs the actual search and hands the results back.

Multi-Step Skills: Doing Real Research

The real power comes when the AI chains multiple skills together to complete a complex task.

You ask: "Summarise the top 3 landlord cases from 2024"

Step 1 → search_cases("landlord disputes", location: "SG")
          returns: [case-001, case-002, case-003, ...]

Step 2 → open_case("case-001")
          returns: full text of case [2024] SGDC 12

Step 3 → open_case("case-002")
          returns: full text of case [2024] SGCA 5

Step 4 → open_case("case-003")
          returns: full text of case [2024] SGHC 88

Step 5 → summarise(case-001 + case-002 + case-003)
          returns: a clear, structured summary

Step 6 → AI replies to you with the finished answer

This is how an AI can handle a task that would normally take a junior researcher an hour — in seconds. It is not guessing or recalling from memory. It is doing real, live work, step by step.

Why This Matters

Without skills, an AI can only draw on what it learned during training. That knowledge has a cutoff date, it has no access to your private systems, and it will sometimes make things up to fill gaps.

With skills, the AI is grounded. It retrieves real, current information rather than recalling something it may have misremembered. The difference is a student guessing at an exam versus a researcher with access to a full library.

What Makes a Good Skill

Not all skills are created equal. The ones that work well tend to share a few traits:

Clear descriptions. The AI picks the right skill based entirely on how it is described. A vague description like “do stuff with cases” will confuse the AI. A specific one like “search legal cases by keyword and location” will not.

One skill, one job. A skill that tries to search and summarise at the same time is harder for the AI to use correctly. The best skills do one thing well.

Clear results. When a skill returns its result, the AI needs to understand it before deciding what to do next. Messy, unstructured output leads to worse answers.

Graceful failures. If something goes wrong — a database is down, a search returns nothing — the skill should say so clearly. The AI can read an error message and adjust. It cannot recover from silence.

In Plain Terms

What you see	What is actually happening
AI answers your question	AI used a search skill, got real data, then wrote a reply based on what it found
AI finds the right case	AI called `search_cases` with your keywords; your database ran the actual query
AI reads a document	AI called `open_case` to load the full text into its working memory
AI chains multiple steps	AI called several skills in sequence, each building on the result of the one before

Skills are what separate a chatbot from a genuine AI assistant. The AI brings the reasoning — deciding what to do and in what order. The skills bring the capability — actually doing it. Together, they form an agent that can handle tasks no single question and answer could.

Curious what all the other AI jargon means — tools, plugins, MCP, hooks? Read our companion post: The AI Agent Terminology Guide.