What is a Prompt?

Introduction

A prompt is any text (or image, audio, file) you give to an LLM as input. The LLM reads it, runs it through billions of learned weights, and produces a continuation. Everything else — system messages, examples, retrieved documents, tool definitions — is just structured prompting.

Beginner analogy: A prompt is the opening line of a conversation. The model's reply is whatever it predicts the rest of the conversation should look like, based on every conversation it has ever seen.

Understanding the topic

Core concepts to understand:

• Prompts are input tokens — chunks of text the model converts to numbers.
• Prompts include the system message, user message, prior turns, and any retrieved context.
• Length matters: prompts must fit inside the model's context window (8k–2M tokens).
• The order of information in a prompt strongly affects the answer (recency & primacy bias).
• A prompt is not just a question — it can be a role, a task, examples, or a full template.

Syntax reference

Visual workflow / architecture:

Prompt = System + Few-Shot Examples + User Message + Retrieved Context

┌─────────────────────────────────────────────────────────┐
│ SYSTEM:  "You are a senior support agent. Be concise."  │
├─────────────────────────────────────────────────────────┤
│ EXAMPLES: Q: ... A: ...   Q: ... A: ...                 │
├─────────────────────────────────────────────────────────┤
│ RETRIEVED CONTEXT: <doc chunks from RAG>                │
├─────────────────────────────────────────────────────────┤
│ USER:    "My subscription was charged twice."           │
└─────────────────────────────────────────────────────────┘

Real-world use

When you chat with ChatGPT, you only see the user message — but behind the scenes OpenAI injects a long system prompt, tool definitions, and memory snippets. Every reply is a continuation of that hidden prompt.

Best practices

• Put the most important instruction at the very top OR the very bottom — those positions get the most attention.
• Keep prompts short enough to fit, long enough to be unambiguous.
• Separate sections with clear delimiters (---, ###, XML tags).

Common mistakes

• Burying critical instructions in the middle of a long prompt.
• Mixing examples and instructions without delimiters.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is a prompt, technically?
• Q2. What parts make up a typical production prompt?
• Q3. Why does the position of an instruction in a prompt matter?
• Q4. What is a context window?