The Parts of an AI Tool

This section follows one question from a web page to an answer and names each part that the question passes through.

Following One Question

An AI tool that a person reaches through a web page is built from several parts that pass work along in order. The example is the Name That Element page of MuggsAI. On that page, a student describes something on a screen, such as a box that pops up over a game, and the tool replies with the name of that element.

First, the student types the description into a box on the page and presses a button. The browser then sends a request to the class site's server. A request is a message that names a server and the resource it asks for, and it can carry data such as the text from a form. Next, the class site's server adds its own instructions to the student's text and sends a second request to Mistral's server, which runs the model. The model has two parts: a large file of numbers called parameters and a smaller program called a runtime. To run, the parameters are loaded into memory, usually the memory on a GPU, and the runtime uses them to predict the answer one token at a time. Finally, the answer travels back as a response, from Mistral's server to the class site's server and then to the browser, which shows it on the page. Figure 1 shows the parts of this chain in order, with the two servers drawn as one box.

The Interface

The interface is the part of a tool that a person sees and touches, which for a web tool is the page in the browser. On the Name That Element page, the interface includes the typing box, the button and the area where the answer appears. MDN describes a user interface as anything that helps a person use a machine, and it lists a web page and an input form as examples. For example, a student who changes the color of the button has changed the interface. The server is part of the tool and not part of the interface. A student never sees the class site's server or Mistral's server, even though both do work on every question. Several MuggsAI pages send their requests to the same server and the same model, and each page still has its own interface.

The System Prompt

A system prompt is a set of instructions that the builder of a tool writes ahead of time and sends to the model with every request. Mistral's documentation describes the system message as an optional message that sets the behavior and context of the assistant. It also says that the system prompt is typically managed by the developer of the tool. In MuggsAI, the server places a system message in front of the student's text each time it sends that text to the model. The Name That Element system prompt tells the model to give names from the class Elements Catalog and not to write code.

What the user types is a prompt, not the system prompt. The student's text travels in a separate message, and that message has the role "user" in the request. The user usually never sees the system prompt, although some tools show it. MuggsOfPrompts, for example, shows its system prompt on the page as read-only text. Because the system prompt shapes how the model responds, two tools that use the same model can behave differently.

Components and the Specification

A component is one part of a system that has one job and can be named and replaced. The interface is a component, because its job is to collect the question and show the answer. The server is a component, because its job is to receive the request, add the system prompt and pass the work along. The model is a component whose job is to predict the answer, and the memory and GPU are the hardware it needs to do that. "The AI" is not one component. It is made of at least four: the interface, the server, the system prompt and the model.

Each component can be changed without rebuilding the others. For example, a builder can rewrite the system prompt and leave the interface and the server unchanged. A specification for a tool describes each component: what it is, what it receives and what it sends on.

Summary

A question typed into a web tool passes through an interface, a request, a server, a system prompt, a model running in memory on a GPU, and a response. Each of these is a component with one job. A specification for a tool names each component and describes its job.

Sources: MDN Web Docs, Client-server overview (2026); MDN Web Docs, UI (2025); Mistral AI, Chat completions and Prompting (n.d.); Hugging Face, Optimizing LLMs for Speed and Memory (n.d.); ggml-org, llama.cpp (n.d.); MuggsAI and MuggsOfPrompts request paths, class site source code (2026).