Introduction

A prompt is the input text given to a language model to guide what it generates — everything from a single question to a detailed set of instructions, examples, and context. Since LLMs work through next token prediction, as covered in earlier topics, the prompt is effectively the starting point that shapes the entire probability distribution the model draws from at every subsequent step, making it the primary lever humans have for directing a model's behavior without changing its underlying weights.

Prompt engineering — the practice of crafting prompts deliberately and skillfully — has become an essential skill for working effectively with generative AI, since the same underlying model can produce dramatically different quality outputs depending on how a request is framed, structured, and detailed.

Why Do Prompts Matter?

Prompts help to:

• Guide a model's behavior and output without requiring any retraining or fine-tuning
• Provide the context, instructions, and constraints a model needs to respond usefully
• Shape tone, format, length, and style of generated content
• Enable a single general-purpose model to perform an enormous range of different tasks
• Reduce ambiguity that could otherwise lead to irrelevant or incorrect responses
• Serve as the most accessible, lowest-cost way to customize model behavior

Where Prompts Fit in the Generation Process

Whiteboard

Loading diagram...

The Anatomy of a Prompt

A prompt can include several distinct components, not always
all present at once:

1. Instruction — what the model should do
   ("Summarize the following article")

2. Context — background information relevant to the task
   ("This article is from a technical blog aimed at developers")

3. Input Data — the actual content to act on
   ("[the article text goes here]")

4. Output Format Guidance — how the response should be structured
   ("Respond in 3 bullet points")

5. Examples (optional) — demonstrations of the desired
   input-output pattern (few-shot prompting)

A Simple Before-and-After Example

Weak Prompt:
"Tell me about dogs."

Strong Prompt:
"You are a veterinary assistant. In 3-4 sentences, explain
the key differences in exercise needs between small and
large dog breeds, written for a first-time dog owner."

The weak prompt leaves the model guessing about scope, tone,
length, and audience. The strong prompt removes that ambiguity,
making a focused, useful response far more likely.

Zero-Shot, One-Shot, and Few-Shot Prompting

Few-Shot Example:

"Classify the sentiment of each review as Positive or Negative.

Review: 'This product exceeded my expectations!' → Positive
Review: 'Completely broke after one use.' → Negative
Review: 'Works okay, nothing special.' → "

The model uses the pattern from the labeled examples to
infer how it should classify the final, unlabeled review.

The Different Roles Within a Prompt (Preview)

Modern LLM APIs typically structure a conversation using distinct roles — System, User, and Assistant — each serving a different purpose in shaping the interaction.

(Each of these roles is explored in full depth in its own dedicated topic.)

Prompt Engineering vs Fine-Tuning

Key Properties of Prompts

• A prompt is the starting input that shapes everything a model generates from that point forward.
• Prompts can combine instructions, context, input data, format guidance, and examples.
• Zero-shot, one-shot, and few-shot prompting offer different levels of example-based guidance.
• Modern LLM interactions are structured around distinct roles — System, User, and Assistant.
• Prompt engineering is typically the fastest, lowest-cost way to adjust model behavior, often tried before fine-tuning.

Where Do Prompts Matter Most?

Advantages

• Requires no model retraining, making it fast and inexpensive to iterate
• Accessible to virtually anyone, without requiring deep machine learning expertise
• Highly flexible — the same model can be redirected toward very different tasks through prompting alone
• Can be combined with other techniques like RAG and fine-tuning for even stronger results
• Provides immediate feedback, since changes can be tested and observed right away

Limitations

• Even well-crafted prompts can't fully eliminate hallucination or reasoning limitations
• Prompt effectiveness can vary between different models, requiring re-tuning when switching models
• Very long, complex prompts consume context window space and increase cost (as covered in earlier topics)
• Prompting alone can't reliably teach a model genuinely new knowledge beyond its training data
• Achieving consistent behavior across many varied inputs can be harder with prompting alone than with fine-tuning

Real-World Examples

Best Practices

• Be specific about the task, audience, tone, and desired format rather than leaving them implicit.
• Provide relevant context and, when helpful, examples to reduce ambiguity.
• Start with prompt engineering before considering fine-tuning, since it's faster and cheaper to iterate on.
• Test prompts across multiple inputs to check for consistency, not just a single favorable example.
• Iterate deliberately — small wording changes can meaningfully affect output quality and consistency.

Interview Tip

A common interview question is:

"What is prompt engineering, and why is it often the first approach tried before fine-tuning a model?"

A strong answer is:

Prompt engineering is the practice of crafting the input given to a language model — including instructions, context, examples, and formatting guidance — to guide its output toward a desired result, without changing any of the model's underlying weights. It's typically tried first because it's fast, inexpensive, and immediately testable, requiring no training process or additional compute, whereas fine-tuning requires curated data and additional training time. Many use cases can be solved effectively through prompt engineering alone, with fine-tuning reserved for situations where consistent, specialized behavior can't be reliably achieved through prompting.

Explaining the speed/cost tradeoff versus fine-tuning makes your answer stronger.

Conclusion

Prompts are the primary interface through which humans direct and shape what a language model generates, making prompt engineering one of the most practical and immediately useful skills for working with generative AI. With this foundation in place, the next topics explore the specific roles — System, User, and Assistant — that structure how prompts are organized within real LLM applications, followed by prompt templates and best practices for context, constraints, and guardrails.