Advanced Techniques
Advanced Prompting Techniques
Once you master the basics, these advanced techniques can help you get even better results from LLMs.
1. Few-Shot Learning
Few-shot learning means providing examples of the task you want the model to perform. The model learns the pattern from your examples and applies it to new inputs.
Zero-Shot (No Examples)
Translate this to Spanish: "Hello, how are you?"One-Shot (One Example)
Translate to Spanish:
Example:
English: Good morning
Spanish: Buenos días
Now translate:
English: Hello, how are you?Few-Shot (Multiple Examples)
Translate to Spanish following these examples:
English: Good morning → Spanish: Buenos días
English: Thank you → Spanish: Gracias
English: See you later → Spanish: Hasta luego
Now translate:
English: Hello, how are you?Expected Output:
Spanish: Hola, ¿cómo estás?When to Use: Few-shot works best for tasks with specific formats, patterns, or styles that might be unusual or domain-specific.
2. Chain-of-Thought (CoT) Prompting
Chain-of-Thought prompting encourages the model to show its reasoning process step-by-step, leading to better answers for complex problems.
Without Chain-of-Thought
A library has 150 books. On Monday, 23 books were checked out and
15 were returned. On Tuesday, 30 books were checked out and 12 were
returned. How many books are now in the library?The model might jump to an answer without showing work.
With Chain-of-Thought
A library has 150 books. On Monday, 23 books were checked out and
15 were returned. On Tuesday, 30 books were checked out and 12 were
returned. How many books are now in the library?
Let's solve this step by step:Expected Output:
Let's solve this step by step:
1. Starting books: 150
2. Monday: 23 checked out, 15 returned
- Net change: 15 - 23 = -8
- Books after Monday: 150 - 8 = 142
3. Tuesday: 30 checked out, 12 returned
- Net change: 12 - 30 = -18
- Books after Tuesday: 142 - 18 = 124
Therefore, there are now 124 books in the library.Few-Shot Chain-of-Thought
Combine few-shot with CoT for complex tasks:
Solve these word problems step by step:
Example 1:
Problem: A store had 50 apples. They sold 20. How many remain?
Solution:
- Starting apples: 50
- Sold: 20
- Remaining: 50 - 20 = 30
Answer: 30 apples
Example 2:
Problem: Tom has 15 books, receives 8 more, then gives away 5. How many does he have?
Solution:
- Starting: 15 books
- Received: +8 → 15 + 8 = 23
- Gave away: -5 → 23 - 5 = 18
Answer: 18 books
Now solve:
A librarian cataloged 45 books on Monday, 33 on Tuesday, and 28 on Wednesday.
On Thursday, 12 books were removed. How many books were cataloged in total?3. Role-Based Prompting
Assign a specific role or persona to the model to influence its responses.
Example: Expert Librarian
You are an expert rare books librarian with 20 years of experience in
manuscript preservation and authentication.
A patron brings in what they claim is a first edition of "Pride and
Prejudice" from 1813. What steps would you take to verify its authenticity?
Provide a systematic approach with specific things to check.Example: Friendly Teacher
You are a friendly and enthusiastic elementary school teacher who loves
making learning fun.
Explain how libraries are organized using the Dewey Decimal System to
a 3rd grade student. Use simple language and examples they can relate to.Example: Technical Expert
You are a senior Python developer with expertise in natural language
processing and machine learning.
Review this code for extracting named entities from text and suggest
improvements for performance and accuracy:
[code here]Why It Works: Roles provide context about expertise level, tone, and approach, helping the model calibrate its response.
4. Structured Output Prompting
Force the model to return data in specific formats like JSON, XML, CSV, or markdown.
JSON Output
Extract book information from this text and return ONLY valid JSON:
"The Great Gatsby by F. Scott Fitzgerald was published in 1925.
To Kill a Mockingbird by Harper Lee came out in 1960."
Format:
{
"books": [
{"title": "...", "author": "...", "year": ...}
]
}
Return only JSON, no additional text.Table/CSV Output
Extract and format as a markdown table:
Text: "London has 9 million people. Tokyo has 14 million. New York has 8 million."
| City | Population (millions) |
|------|----------------------|Bullet List Output
Summarize the key points from this article as a bullet list:
- Each point should be one sentence
- Use past tense
- Focus on facts only
- Maximum 5 points5. Constraint-Based Prompting
Add specific constraints to guide the model’s behavior.
Length Constraints
Explain quantum computing.
Constraints:
- Exactly 50 words
- Use no technical jargon
- Include one analogyContent Constraints
Write a product description for this book.
Requirements:
- Do NOT mention price
- Do NOT use superlatives (best, greatest, etc.)
- Do NOT make claims about reviews or ratings
- Focus only on content and target audience
- Maximum 100 wordsStyle Constraints
Summarize this research paper.
Style requirements:
- Use present tense only
- Write in third person
- Use passive voice for methods
- No contractions
- Academic tone6. Multi-Step Prompting
Break complex tasks into sequential steps.
Example: Research Summary Pipeline
Step 1: Extract key information
From this research paper, extract:
1. Main research question
2. Methodology used
3. Key findings (top 3)
4. Limitations mentioned
[paper text]Step 2: Synthesize findings
Using this extracted information:
[output from Step 1]
Write a 150-word abstract suitable for a conference submission.Step 3: Generate questions
Based on this abstract:
[output from Step 2]
Generate 5 critical questions that reviewers might ask.Pro Tip: For multi-step tasks, use separate prompts rather than asking the model to do everything at once. This improves accuracy and gives you checkpoints to verify.
7. Negative Prompting
Tell the model what NOT to do.
Example
Explain machine learning to a business executive.
DO:
- Use business-relevant examples
- Focus on practical applications
- Keep it under 200 words
DO NOT:
- Use mathematical equations
- Mention specific algorithms
- Get into technical implementation details
- Use computer science jargonWhy It Works: Explicitly stating what to avoid prevents common errors and keeps the model focused.
8. Iterative Refinement
Use the model’s output as input for improvement.
Step 1: Initial Prompt
Write a brief description of what vector databases are.Step 2: Refine
Here's a description of vector databases:
[previous output]
Improve this description by:
1. Adding a concrete real-world example
2. Explaining why they're better than traditional databases for certain tasks
3. Keeping it under 150 wordsStep 3: Further Refine
Here's the improved description:
[previous output]
Make it more accessible to someone with no database experience.
Replace technical terms with analogies.9. Template-Based Prompting
Create reusable templates for common tasks.
Book Review Template
Write a book review following this template:
Book: [TITLE] by [AUTHOR]
Genre: [GENRE]
Structure:
1. Opening Hook: One engaging sentence about the book
2. Summary: 2-3 sentence plot summary (no spoilers)
3. What Works: 2 strengths of the book
4. What Could Be Better: 1 constructive criticism
5. Recommendation: Who would enjoy this book?
6. Rating: X/5 stars with brief justification
Total length: 200-250 words
Tone: Balanced, thoughtful, accessibleData Extraction Template
Extract data from text using this template:
INPUT: [text]
EXTRACT:
- Entities: [list of people, places, organizations]
- Dates: [list of dates mentioned]
- Key Numbers: [important numerical data]
- Topics: [main themes, max 3]
FORMAT: JSON10. Comparative Prompting
Ask the model to compare and contrast for deeper analysis.
Example
Compare and contrast these two books based on the following text:
[text mentioning multiple books]
Create a comparison table with these dimensions:
- Publication era
- Language/Culture
- Genre
- Estimated complexity level
- Target audience
Present as markdown table.Combining Techniques
The most powerful prompts often combine multiple techniques:
Example: Complex Extraction Task
[ROLE] You are a data analyst specializing in literary metadata.
[TASK] Extract book information from the text below and analyze patterns.
[FEW-SHOT EXAMPLE]
Input: "Orwell's 1984 was published in 1949."
Output: {"title": "1984", "author": "Orwell", "year": 1949}
[CONSTRAINT] Only include books published between 1800-2000.
[OUTPUT FORMAT]
1. JSON array of books
2. Summary paragraph identifying patterns (languages, time periods, genres)
[CHAIN OF THOUGHT] Show your analysis reasoning.
[TEXT]
{insert sample text}Advanced Tips
1. Use Delimiters
Separate different parts of your prompt clearly:
###INSTRUCTION###
Summarize the following text
###CONSTRAINTS###
- Maximum 100 words
- Focus on key findings
###TEXT###
[your text here]
###OUTPUT###2. Specify Tone and Audience
Write this for: [audience description]
Tone: [formal/casual/technical/etc.]
Reading level: [grade level or expertise]3. Request Self-Critique
[task]
After providing your answer, critique it and suggest one improvement.4. Use System Messages (When Available)
Some interfaces let you set a system message that persists:
System: You are a meticulous fact-checker. Always cite sources when possible
and express uncertainty when appropriate.
User: Tell me about the history of libraries.Practice Exercise
Try improving this weak prompt using techniques from this section:
Weak Prompt:
Tell me about the booksYour Improved Prompt Should Include:
- Clear context
- Specific instruction
- Output format
- At least one advanced technique
Improved Prompt:
[ROLE] You are a literary analyst specializing in world literature.
[CONTEXT] I have a text describing various books from different cultures
and time periods.
[TASK] Analyze the following text and extract book information.
[OUTPUT FORMAT]
1. Create a JSON array of books with: title, author, year, language, genre
2. Write a 100-word analysis of geographic and temporal distribution
3. Identify the 3 most culturally significant works and explain why
[CONSTRAINTS]
- Only include books explicitly mentioned with clear publication dates
- Use ISO 639-1 language codes (en, es, ja, etc.)
- Analysis should be objective and data-driven
[TEXT]
{insert your text here}Next: Hands-On Practice
Ready to put these techniques into action?
Mastering These Techniques: Don’t try to use all techniques at once. Pick 1-2 that fit your task and master them first. You can always combine more as you gain experience.