What Do AI Buzzwords Like LLM, Agent, and Skill Really Mean?

LLM – foundation of all capabilities

LLM (Large Language Model) refers to a massive language‑prediction system built on the Transformer architecture originally proposed by Google and popularized by OpenAI. After extensive training, the model predicts the most likely next token, enabling writing, question answering, data summarization, and code generation.

Token – basic processing unit

Input text is first passed through a tokenizer, which splits the text into tokens and maps each token to a numeric ID. The model processes numbers; its output numbers are later converted back to text.

A token is not strictly a character or a word; for example, a Chinese character pair may form one token, while an English word can be split into several tokens. Tokens are the model’s information granules, generated sequentially (e.g., producing “should” then “be” until the answer is complete).

Context – information visible to the model

During a conversation, the model “remembers” previous dialogue because the current request includes a context: the question, dialogue history, system instructions, uploaded files, tool results, etc. The quality of the answer depends on whether this context is complete, accurate, and relevant.

The context window defines the maximum number of tokens the model can ingest at once. Exceeding it requires techniques such as Retrieval‑Augmented Generation (RAG) to fetch the most relevant excerpts before feeding them to the model.

Prompt – task description sent to the model

A prompt is the specific instruction given to the model, e.g., “Help me pick a melon.” If the prompt is vague, the model may not know which type of melon or which criteria matter.

Clear prompts specify the goal, constraints, and desired output format. While modern models can often infer intent from ambiguous prompts, precise prompts still significantly improve output quality for complex tasks.

Tool – external capability accessed by the model

Without tools, a model cannot obtain real‑time information; it can only provide knowledge up to its training cutoff. Tools enable the model to query the web, call maps, read databases, manipulate files, execute code, or control software interfaces.

Typical workflow: the model decides which tool to use, the tool performs the action, and the platform orchestrates the interaction.

Model: selects tool and aggregates results. Tool: executes the query. Platform: connects the components.

MCP – unified standard for tool integration

When many tools exist, each platform may require a different integration code. MCP (Model Context Protocol) proposes a single interface so that a tool can be used across OpenAI, Google, Anthropic, and other platforms without rewriting adapters.

Agent – system that plans and executes tasks

Even with tools and MCP, a model alone cannot orchestrate complex workflows. An agent adds a planner that breaks a goal into steps, calls tools, checks results, and iterates until the task is complete.

Example: finding the cheapest nearby watermelon requires understanding the goal, decomposing steps (search price, compare, plan route), invoking tools, and presenting the final answer. Products such as Claude Code, Codex, and Gemini demonstrate this capability.

Agent Skill – reusable instruction package for agents

When the same type of task is performed repeatedly, writing the full prompt each time is cumbersome. An Agent Skill encodes the workflow, rules, tool usage, and precautions for a specific task, acting like a long‑term SOP. It does not replace the model’s abilities but makes the agent’s behavior more stable and aligned with user expectations.

Term relationships

LLM – base model that generates language.

Token – basic unit of text processed by the model.

Context – all information the model can see in a request.

Prompt – specific instruction given to the model.

Tool – external capability the model can invoke.

MCP – open standard that unifies tool‑model integration.

Agent – system that plans, calls tools, and iterates to achieve a goal.

Agent Skill – reusable knowledge package that defines how an agent performs a particular task.

Understanding this chain—LLM as the base, Token and Context defining input/output boundaries, Prompt as the entry point, Tool and MCP providing external capabilities, Agent orchestrating them, and Skill stabilizing Agent behavior—clarifies how to work with AI products.