P1-12.2 Instructions, Context, and Examples¶

Section ID: P1-12.2 Version: v2026.07.20

Section 12.1 described a prompt as a way of arranging task conditions inside the current input. A prompt does not retrain the model. Instead, it places instructions, context, examples, constraints, and output format inside the input that the model can use during inference.

This section separates the three most basic parts of that structure:

instruction:
what is the model being asked to do?

context:
what should it refer to while doing that?

example:
what desired input-output pattern is being shown?

Once these three are distinguished, a prompt becomes easier to read not as a polished sentence, but as an input structure that arranges task conditions by role.

Part 1 introduces the basic intuition of instructions, context, examples, few-shot prompting, and structured prompting here. Section 12.1 first explained what prompts specify. This section separates the three most central elements. Evaluation criteria and limits are handled separately in 12.3.

This section does not cover advanced prompt-engineering techniques in detail. Techniques such as Chain-of-thought, self-consistency, and automatic prompt optimization are not explained deeply here. They return later in the practical context of Part 5 P5-9.3.

These terms can all sound like decorative ways of making a prompt longer. A quick distinction helps:

The minimum distinction to keep is:

• the instruction is the task
• context is the reference information
• an example is the demonstration
• few-shot prompting means a prompt that includes examples

This section also does not cover RAG, tool use, or agents. RAG returns in P1-13.3 and P1-13.4. Tool use and agents return in P1-14.2 and P1-14.3. The only goal here is to learn the basic components that should be separated first when giving requests to an LLM.

This section also does not ask whether a prompt guarantees a good result. That question is deferred to 12.3 through the lenses of limits, evaluation, and reproducibility.

Separating Instructions, Context, and Examples¶

• Distinguish instruction, context, and example.
• Explain how the three can be arranged separately inside one prompt.
• Understand that examples do not retrain the model, but show patterns inside the current input.
• Understand that even with context, factual verification is not automatically complete.
• Prepare naturally for the limits and evaluation criteria handled in 12.3.

Three Standards¶

The Instruction Specifies the Task¶

An instruction is the part that tells the model what to do.

summarize this
compare these two things
rewrite this for beginners
organize it as a table
mark weakly supported statements as needs verification

If the instruction is vague, the model interprets the task broadly.

vague instruction:
tell me about AI

clearer instruction:
explain the relationship among AI, machine learning, deep learning, and generative AI to a reader reviewing an introductory AI course

Both prompts are natural-language requests. But the second one specifies the task target, the reader, and the concepts to compare much more clearly.

An instruction is not better just because it is longer. A good instruction narrows the task.

In writing this book, it is especially important to include the target reader and the review standard inside the instruction.

explain this to readers who studied introductory AI a long time ago but forgot much of it
separate personal interpretation from standard explanation
do not write unsupported forecasts

That kind of instruction tells the model not only what to write, but also by what standard to write it.

Context Provides the Information to Refer To¶

Context is the information the model should use while carrying out the instruction. This can include background, source text, purpose, earlier decisions, reader conditions, or excluded scope.

For example, this prompt has only an instruction:

rewrite this sentence more simply

The model still does not know what sentence it should work on. Once context is added, the task becomes possible:

task:
explain the following sentence more simply for a reader reviewing introductory AI

original sentence:
Transformers compute relationships among input tokens using self-attention.

context:
this document prioritizes conceptual recovery over mathematical detail

Here, task is the instruction. The original sentence and the contextual note are reference material.

Context does not mean only general background description. The following can all count as context:

When giving context, an important distinction is:

what is left for the model to guess
versus
what is explicitly provided as input

Compare:

prompt that leaves guessing:
write the next section

prompt with context:
the previous section explained what prompts specify
this section distinguishes instruction, context, and example
the next section will cover limits and evaluation
so do not go deeply into those limits here

The second prompt tells the model the chapter position and section boundary. This matters especially in long documents.

Examples Show the Desired Pattern¶

An example is the part that demonstrates the input-output pattern you want. Brown and colleagues described the few-shot setting in GPT-3 as including a few task demonstrations inside the input without additional fine-tuning.

At an introductory level:

instruction:
write terms in Korean with the English original

example:
input: AI
output: 인공지능(artificial intelligence)

input: ML
output: 머신러닝(machine learning)

input: inference
output:

The example shows the model the pattern of the desired answer. That increases the chance that it continues with something like 모델 실행(inference) or, depending on context, another context-appropriate bilingual form.

Examples can also show not only answer format, but judgment criteria.

instruction:
review the following sentence and separate standard explanation from personal interpretation

example:
sentence: Deep learning is a direct copy of the human brain.
review:
- standard explanation: artificial neural networks were inspired by biological neural networks, but are not literal copies of the human brain
- personal interpretation: it is acceptable as a loose analogy for intuition
- judgment: risk of exaggeration

sentence: A prompt does not change model weights.
review:

This example does more than show a formatting pattern. It also shows the review frame itself: standard explanation, personal interpretation, and judgment.

Examples still need caution.

Examples are not permanent knowledge updates. They are patterns supplied for the current input.

Once the Three Parts Are Separated, the Prompt Becomes Readable¶

The first step to writing better prompts is not to craft stylish sentences. It is to separate roles inside the input.

task:
rewrite the following paragraph for readers reviewing introductory AI

context:
- this document is study material for recovering concepts learned long ago
- personal interpretation must be separated from standard explanation
- this section does not cover RAG or agents

original text:
prompts are a means to understand how LLMs behave

example:
good wording:
prompts do not reveal the inside of an LLM directly,
but they let us observe how outputs change under different input conditions

wording to avoid:
by looking at prompts, we can know the internal principle of an LLM

output format:
- revised sentence
- reason for revision
- points that need verification

This prompt separates instruction, context, source text, example, and output format. That makes it easier for a reviewer to see what is being requested, and more likely that the model can distinguish the roles played by each part.

Of course, structuring a prompt this way does not guarantee a correct answer. A prompt organizes input conditions. It does not replace fact checking or quality evaluation. Those limits return in 12.3.

A Basic Template That Works Well for Long Learning Documents¶

For long learning documents, the following template is often useful:

task:
what is being requested?

target:
what original text or material is being reviewed or written?

context:
what are the document position, reader, earlier decisions, and excluded scope?

criteria:
what are the standards for evidence, terminology, bilingual notation, section boundaries, and avoiding exaggeration?

example:
what short example shows the desired format or judgment style?

output:
in what shape should the result be returned?

This template does not fit every situation exactly. But when a long book is written with AI tools, it helps structure the request and leaves clearer points for later human review.

Checklist¶

• I can explain that an instruction specifies the task the model should perform.
• I can explain that context provides background information and materials needed for the task.
• I can explain that examples show the desired input-output pattern.
• I can explain that examples do not change model weights the way fine-tuning does.
• I can read a long prompt by separating task, context, source text, example, and output format.
• I can explain that prompt structuring does not replace fact verification.
• I can read a long prompt by separating it into what to do, what to refer to, and what pattern to follow.
• I can explain that few-shot prompting does not retrain the model, but only presents a pattern inside the current input.

That shifts the problem away from making the prompt longer and toward making its roles readable and checkable.

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.