P1-12.1 What Does a Prompt Specify?¶

Section ID: P1-12.1 Version: v2026.07.20

Chapter 11 explained where LLMs came from. Language models began from the problem of next-token probability, and Transformers and pretraining became key foundations of modern LLMs.

This section now turns to how users interact with LLMs. The starting point is the prompt.

The central question is:

what kind of input does a prompt provide to the model?

The key perspective is:

a prompt is not training data that retrains the model;
it is a way of placing instructions, context, examples, constraints, and output format inside the current input so the model can use them while generating the response

Part 1 introduces the basic distinctions among prompts, instructions, context, examples, constraints, output format, and prompting versus fine-tuning here. Section 11.3 already distinguished in-context learning from pretraining. This section continues from there and organizes the question:

what can a user specify through input?

Prompts do not reveal the internal mechanics of an LLM directly. But they do make it possible to observe how output changes when the same model receives different input conditions. So a prompt is not proof of the model’s inner structure. It is better treated as an external entry point for observing inference behavior.

This section does not cover a catalog of prompt-engineering techniques. Chain-of-thought, few-shot example design, and role prompting return in P1-12.2. Evaluation prompts and limit analysis return in P1-12.3.

These terms can all sound like similar input ingredients at first. A quick distinction helps:

The minimum distinction to keep is:

• the prompt is the whole input
• the instruction is the task
• context is background
• an example shows a pattern
• constraints define boundaries

This section focuses only on what prompts specify:

A prompt is not a magic spell. Even with the same model, different input arrangements can lead to different outputs, but prompts do not change model weights.

The focus here is to understand prompts as input structure. Example placement and evaluation criteria are deferred to 12.2 and 12.3.

Input Conditions Specified by a Prompt¶

• Understand a prompt as natural-language input given to an LLM.
• Explain that prompts can specify instructions, context, examples, constraints, and output format.
• Understand that prompts let us observe how input conditions affect generation.
• Distinguish prompting from fine-tuning.
• Understand that prompts do not guarantee factual verification or truth.
• Treat good prompting not as clever wording, but as clear input specification for a task.

Three Standards¶

A Prompt Is Input for Building the Current Response¶

An LLM reads input tokens and generates the next tokens. A user’s question, command, document, example, or dialogue history all become part of the input the model can use while producing the next output.

That is why prompts help us understand LLM behavior. If the same topic is asked with different target readers, scope, examples, or forbidden conditions, the output changes. That does not explain the internal circuitry of the model, but it does let us observe that an LLM is sensitive to the current input context.

For example:

explain the history of AI to a reader who is seeing it for the first time

This prompt already specifies at least three things:

If the prompt becomes more explicit:

explain the history of AI to a first-time reader
distinguish the flow of rule-based AI, machine learning, deep learning, and LLMs
keep each stage within three sentences
avoid exaggerated phrasing

Now the prompt specifies not only the topic, but also scope, structure, length, and tone. The model uses these as part of the input when generating the response.

Prompting Is Not Learning, but Condition Specification¶

Giving a prompt does not mean the model has learned something new. As Section 11.3 explained, in-context learning is safer to understand as a change in current output behavior caused by instructions and examples inside the input context, without updating model weights.

fine-tuning:
update model weights using training data

prompting:
place task conditions inside the current input to guide the output

The GPT-3 paper by Brown and colleagues explains zero-shot, one-shot, and few-shot settings while keeping the model fixed and providing task and demonstration through textual interaction rather than gradient updates or fine-tuning.

So the safer distinction is:

changing the model itself:
pretraining, training, fine-tuning

changing the conditions of the current response:
prompts, examples, dialogue history, provided documents

If this distinction is lost, people can misunderstand prompts as if they make the model learn the fact permanently, or as if something told once will always remain inside the model afterward.

A Prompt Can Contain Both Instruction and Context¶

Prompts usually contain both instruction and context.

instruction:
what is the model being asked to do?

context:
what information should the model refer to while doing it?

For example:

rewrite the following sentence as an introductory explanation

sentence:
Transformers compute sequence representations using self-attention.

Here, rewrite the following sentence as an introductory explanation is the instruction. Transformers compute... is the target text and also the local context.

As prompts become longer, instruction and context can easily blur together. That is why it is often safer to separate them explicitly:

task:
rewrite the following sentence for an introductory reader

original sentence:
Transformers compute sequence representations using self-attention.

conditions:
- keep it within three sentences
- use only one analogy
- avoid phrases like “understands like a person”

That makes it more likely that the model can distinguish the roles played by different parts of the input. But even this is not a guarantee. A prompt is a way to express conditions clearly, not a device that guarantees correctness automatically.

Examples Show a Desired Output Pattern¶

An example is input that shows the model the pattern you want.

input: AI
output: artificial intelligence

input: ML
output: machine learning

input: LLM
output:

Given these examples, the model becomes more likely to continue with large language model.

This is the basic intuition of few-shot prompting: place several input-output demonstrations inside the prompt so the model can infer the current task’s form and intention.

Examples can be powerful, but they also carry risks.

bad example:
can reinforce the wrong pattern

biased example:
can narrow the direction of output too much

overly long example:
can push important instructions out of the effective context window

Examples are not permanent training data. They are current patterns to be referenced inside the input. That is why the quality of examples can directly affect the quality of outputs.

Constraints Narrow the Range of Results¶

Prompts can also give constraints that narrow the allowed range of outputs.

explain it in under 200 characters
organize it as a table
if there is no verifiable evidence, mark it as needs verification
do not use difficult equations for first-time readers
write Korean terms with the English original in parentheses

Constraints tell the model the boundaries of the task. In writing this book, constraints matter a great deal.

For example, these two prompts are likely to produce different results:

prompt A:
explain LLMs

prompt B:
explain LLMs to a reader revisiting an introductory AI course
cover only the relation among Transformers, pretraining, and prompts
leave RAG and agents to the next chapter
do not include unsupported forecasts

Prompt B specifies the task, reader, scope, excluded material, and verification standard together. The reason it is more useful is not that it is longer, but that the necessary conditions are clearer.

Prompts Do Not Guarantee Factuality¶

A better prompt can raise the chance of getting a more useful answer. But prompts alone cannot guarantee factuality, evidence, or safety.

For example, this is dangerous:

write this claim so it sounds convincing and factual

That kind of prompt can encourage unsupported statements to appear persuasive. In this book, the preference is the opposite:

write only verified content as factual
if the evidence is weak, label it as needs verification
separate personal interpretation from standard explanation

Prompts can guide output, but they do not replace verification responsibility. In a learning-oriented book, evidence review matters more than prompt wording.

Checklist¶

• I can explain a prompt as the current input given to an LLM.
• I can distinguish instruction from context.
• I can explain that examples show output patterns.
• I can explain that constraints narrow the range of the result.
• I can explain that output format can also be specified in the prompt.
• I can explain that prompts do not directly reveal the inside of the LLM, but do let us observe changes in output under different input conditions.
• I can explain the difference between prompting and fine-tuning.
• I can explain in-context learning as a behavior change inside the current input context without weight updates.
• I can explain that prompts do not automatically guarantee factuality, evidence, or safety.
• I can explain prompts not as tips for saying things nicely but through the structure of current input, task conditions, no weight update, and separate verification responsibility.
• I can distinguish where prompting, fine-tuning, and in-context learning diverge.

Sources and Further Reading¶

• Tom B. Brown et al., Language Models are Few-Shot Learners, arXiv, 2020, accessed 2026-06-23.
• Long Ouyang et al., Training language models to follow instructions with human feedback, arXiv, 2022, accessed 2026-06-23.
• Pranab Sahoo et al., A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications, arXiv, 2024, accessed 2026-06-23.
• OpenAI, Prompt engineering, OpenAI API documentation, accessed 2026-07-19.