Prompting

We will focus on text based prompting

Key resources

• The Prompt Report: A Systematic Survey of Prompting Techniques
• Prompt engineering using Instructor

Some definitions first

What’s a prompt?

• “A prompt is an input to a Generative AI model, that is used to guide its output”
• “Prompts may consist of text, image, sound, or other media”

Prompting is for in-context learning (ICL)

• “the ability of GenAIs to learn skills and tasks by providing them with exemplars and or relevant instructions within the prompt”

• No retraining of models (no model weight updates)

What’s a prompt template?

• “A prompt template is a function that contains one or more variables which will be replaced by some media (usually text) to create a prompt.”
• “a prompt template becomes a prompt when input is inserted into it.”

What’s a prompt template?

• “a prompt template becomes a prompt when input is inserted into it.”

• Template: “Write a poem about {{TOPIC}}”

  • Prompt: “Write a poem about HCAI”

Can have multiple variables

• “Write a {{POEM FORM}} about {{TOPIC}}”
  • Write a sonnet about HCAI

What’s a prompt template?

• “a prompt template becomes a prompt when input is inserted into it.”

• Prompt or prompt template? 🧐

  • Write a {{POEM FORM}} about HCAI?

Prompt chain

• “two or more prompt templates used in succession”
• “The output of the prompt generated by the first prompt template is used to parameterize the second template, continuing until all templates are exhausted”

Prompt engineering

• “the iterative process of developing a prompt by modifying or changing the prompting technique that you are using”

Prompt components

• Directives
• Examples/Shots/Exemplars
• Output Formatting
• Style
• Role
• Context
  • Additional Information

Directives — what’s to be done?

• Explicit
  • “Write a poem”
• Implicit
  • “Good morning: buen día
  • Good night: ”

Examples/Exemplars/Shots

Demonstrations that guide the GenAI to accomplish a task

1-shot

• “Good morning: buen día
• Good night: ”

Few-shot

2+2: four
4+5: nine
8+0:

How many shots?

System prompt from Anthropic

• Todo
  • Take a look at the system prompt for Claude Opus 4.5
    • <refusal_handling>
    • <user_wellbeing>

Text based prompting techniques

Text based prompting techniques — lots of options!

Text based prompting techniques

• Zero-shot
• Few-shot
• Chain of thought (CoT)

Zero-shot prompting techniques

• Role prompting
• Style prompting
• Emotion prompting
• S2A (System 2 Attention)
• Rephrase and Respond (RaR)
• Self-Ask

Role prompting — assign a role

• “You are a great poet”
• “You are an expert financial advisor”

Style prompting — desired style

• “Be concise”
• “Respond in songs”

Emotion prompting — using phrases of “psychological relevance to human”

• “This is very important to my career”

Emotion prompting — using phrases of “psychological relevance to human”

• “Are you sure that’s your final answer? Believe in your abilities and strive for excellence. Your hard work will yield remarkable results”
  • Social Cognitive Theory

S2A (System 2 Attention)

• Two steps
  • Ask to rewrite the prompt to remove any unnecessary information
  • Use the rewritten prompt for the answer

Rephrase and Respond (RaR)

• Clarify ambiguous information in the prompt
  • “Rephrase and expand the question”

Self-Ask

• Decide if follow-up questions are required
  • generate the follow-up questions
  • answer the follow-up questions
  • answer the main query

Text based prompting techniques

• Zero-shot
• Few-shot
• Chain of thought (CoT)

Chain of thought (CoT)

• Explain the reasoning process before answering question
• CoT can be Zero-shot or few-shot

Zero-shot CoT

• Step-back prompting
• Analogical prompting

Step-back prompting

• Abstraction
  • Ask the LLM a generic, higher-level concept
    • Step back question
• Reasoning
  • “Ask the LLM the original question, given its answer to the abstract question”
• Example

Analogical prompting

• “generate examples that are relevant to the problem”
• Example

Few-shot CoT

• Contrastive few-shot CoT
  • Exemplars with both incorrect and correct for the CoT prompt
    • How to and how not to reason
• Example

Meta prompting

Todo

• Create a meta prompt and run it in GPT or Claude

Output formatting

• We often want structured output

• When you are using LLMs as a pipeline or infrastructure for further processing, you want consistency

Free form text output is not fun to parse

We want structured output (using Instructor)

import instructor

class Response(BaseModel):
    message: str

r = client.responses.create(
    input="Write a haiku about IST 597 (HCAI)",
    response_model=Response,
)

Define the structure of response from LLMs

import instructor

class Response(BaseModel):
    message: str

r = client.responses.create(
    input="Write a haiku about IST 597 (HCAI)",
    response_model=Response,
)

response_model: expected structure

Different data types

class User(BaseModel):
    name: str
    age: int

Todo: Using different data types

• Create a new response model that only includes integers (int)
• Use it in a prompt that returns an integer
  • What happens if your prompt does not return an integer?

Todo: Anthropic prompt library

• Check the prompt library
• Pick two of your favorite prompts
  • Do they actually work?