From LLM to Agent Skill

A linked walkthrough of the concepts LLM / Token / Context / Context Window / Prompt / Tool / MCP / Agent / Agent Skill

Model Layer

This line starts with LLM, Token, Context, and Prompt.

• LLM is the underlying model. At its core it is a parameterized prediction system
• Token is the basic unit the model processes. It is not the same as a single natural-language character
• Context is the input material the model can actually see in the current inference
• Context Window is the capacity limit of that material

Prompt is what packs the task, role, and other restrictions into the current context. System Prompt is where the developer sets preset boundaries, while User Prompt carries the user's current task.

Capability Layer

When the model can only continue text, it is only an answer system, not an execution system.

• Tool: lets the model call external functions or services
• MCP: standardizes tool and context integration, reducing the cost of every client having to write its own plugin. It is usually implemented by external service providers and installed by users for model use
• Agent: not just a one-shot answer, but an abstract intelligent bundle that can make multi-step decisions around a goal, call tools, and collect results
• Agent Skill: packages high-frequency task flows, constraints, and reference materials into a reusable entry point

The easiest things to confuse here are MCP, Agent, and Skill. MCP solves integration, not intelligence. Agent solves execution. Skill solves reuse. They are not synonyms. They just sit on different layers of the same chain.

Layered Troubleshooting

When using AI, treating every problem as a prompt problem is usually wrong:

• Poor output quality is usually caused by Context or Prompt design
• Tool integration problems make it impossible to read external information
• Lack of closure at the Agent layer causes the model to answer but fail to complete tasks stably
• Missing Skill means the same kind of task has to be re-explained every time

Mapping the problem back to the right layer makes troubleshooting faster in development.

Concept Layers

• Model understanding: Token, Context, Context Window, Prompt
• What the model can do: Tool, MCP
• Stable system execution and reuse: Agent, Agent Skill

This chain is pushing the model from a single "text generation" system toward a system shape that is integrable, executable, and reusable.