Advanced Prompting Techniques

Beyond basic prompting, several advanced techniques can dramatically improve LLM output quality, accuracy, and reliability.

Zero-Shot Prompting

Asking the model to perform a task without any examples:

Classify the sentiment of this review as Positive, Negative, or Neutral:

"The food was decent but the service was incredibly slow. Won't be returning."

Sentiment:

Zero-shot works well for straightforward tasks where the model has seen similar patterns in training.

Few-Shot Prompting

Providing examples of the desired input-output pattern before asking the actual question:

Classify the sentiment:

Review: "Absolutely loved it! Best purchase ever."
Sentiment: Positive

Review: "It's okay, nothing special."
Sentiment: Neutral

Review: "Terrible quality, broke after one day."
Sentiment: Negative

Review: "The food was decent but the service was incredibly slow."
Sentiment:

How Many Examples?

Examples	Name	Best For
0	Zero-shot	Simple, well-defined tasks
1-3	Few-shot	Format and pattern matching
4-8	Many-shot	Complex classification, nuanced tasks

Tips for few-shot examples:

Use diverse, representative examples
Cover edge cases
Keep examples consistent in format
Order can matter — put the most relevant examples last

Chain-of-Thought (CoT) Prompting

Ask the model to show its reasoning step by step. This dramatically improves accuracy on reasoning tasks.

Without CoT:

Q: A store has 15 apples. 8 are sold in the morning and 3 more arrive
   in the afternoon. How many apples are there at the end of the day?
A: 10

With CoT:

Q: A store has 15 apples. 8 are sold in the morning and 3 more arrive
   in the afternoon. How many apples are there at the end of the day?

Think step by step:

A: Let me work through this:
   1. Starting apples: 15
   2. After morning sales: 15 - 8 = 7
   3. After afternoon delivery: 7 + 3 = 10

   The store has 10 apples at the end of the day.

Simple CoT Trigger Phrases

"Think step by step"
"Let's work through this carefully"
"Show your reasoning"
"Break this down into steps"

Zero-Shot CoT

Simply adding "Let's think step by step" to any prompt:

"What is 17 × 23? Let's think step by step."

→ 17 × 23
  = 17 × 20 + 17 × 3
  = 340 + 51
  = 391

This single phrase improves math, logic, and reasoning tasks significantly.

Role Prompting

Assigning a specific persona or expertise to the model:

You are a senior security engineer at a FAANG company with 15 years
of experience in application security. You specialize in finding
vulnerabilities in web applications.

Review this authentication code and identify all security issues:

[code]

Effective Role Prompts

Role	Best For
"Expert in X"	Deep technical analysis
"Teacher explaining to beginners"	Simplified explanations
"Devil's advocate"	Finding counterarguments
"Code reviewer at Google"	High-quality code review
"Technical writer for API docs"	Documentation

Self-Consistency

Generate multiple responses to the same prompt and take the majority answer. Useful for reasoning tasks where the model might err.

# Run the same prompt 5 times:
Prompt: "Is 17 a prime number? Think step by step."

Response 1: Yes (correct reasoning)
Response 2: Yes (correct reasoning)
Response 3: No (flawed reasoning)    ← outlier
Response 4: Yes (correct reasoning)
Response 5: Yes (correct reasoning)

Majority vote: Yes ✓

In practice, you can do this programmatically by setting temperature > 0 and calling the API multiple times.

Generated Knowledge Prompting

Ask the model to generate relevant knowledge before answering:

Step 1: Generate relevant facts about electric vehicles and
        environmental impact.

Step 2: Using those facts, answer: "Are electric vehicles truly
        better for the environment than gas cars?"

This two-step approach produces more informed, nuanced responses.

Prompt Chaining

Breaking a complex task into a sequence of simpler prompts, where each output feeds into the next:

Prompt 1: "Extract all the key claims from this article"
           → [List of claims]

Prompt 2: "For each claim, identify what evidence would be needed
            to verify it"
           → [Evidence requirements]

Prompt 3: "Rate each claim's credibility based on the evidence
            requirements. Use: Verified / Likely / Uncertain / Unlikely"
           → [Final analysis]

When to Chain vs Single Prompt

Use Chaining	Use Single Prompt
Complex multi-step tasks	Simple, focused tasks
Need intermediate validation	One-shot generation
Each step requires different expertise	Consistent format needed
Output quality is critical	Speed matters more

Negative Prompting

Telling the model what NOT to do:

Explain quantum computing to a high school student.

DO NOT:
- Use jargon without defining it
- Include mathematical formulas
- Make it longer than 200 words
- Use condescending language like "simply put" or "it's easy"

ReAct (Reasoning + Acting)

Combining reasoning with actions in a structured format:

Question: What is the population of the capital of Australia?

Thought 1: I need to find the capital of Australia first.
Action 1: The capital of Australia is Canberra.
Thought 2: Now I need to find the population of Canberra.
Action 2: Canberra has a population of approximately 467,000.
Answer: The population of Canberra, the capital of Australia,
        is approximately 467,000.

This pattern is foundational for AI agents (covered in Module 7).

Comparison Table

Technique	Best For	Complexity
Zero-shot	Simple tasks	Low
Few-shot	Pattern matching, classification	Low
Chain-of-thought	Math, logic, reasoning	Medium
Role prompting	Domain expertise	Low
Self-consistency	High-accuracy answers	Medium
Prompt chaining	Complex multi-step tasks	High
ReAct	Tasks requiring tools/actions	High

Practice Exercises

Use chain-of-thought to solve: "If a train travels at 60 mph for 2.5 hours, then 80 mph for 1.5 hours, what's the total distance?"
Create a few-shot prompt to classify customer support tickets into categories
Use role prompting to get a code review from different perspectives (security, performance, readability)

These techniques form the foundation for building reliable AI applications. The next topic covers reusable prompt patterns and frameworks.