Prompt Engineering Basics

Comprehensive guide to prompt engineering fundamentals for working effectively with generative AI models like ChatGPT, Claude, Gemini, Cohere Chat, and other large language models. Master essential and advanced techniques to maximize AI tool potential in your daily tasks.

When to Use

Use this skill when you need to:

•Craft effective prompts for any generative AI model or chatbot
•Improve AI output quality through better instruction design
•Learn prompt engineering fundamentals as a beginner
•Apply structured prompting techniques (zero-shot, few-shot, chain-of-thought, self-criticism, decomposition, ensembling)
•Debug poor AI responses by understanding what went wrong
•Design AI-powered workflows requiring consistent, high-quality outputs
•Evaluate prompt performance across different models

Not recommended for: Advanced prompt optimization, model fine-tuning, RAG implementation, or AI safety research (see references).

Instructions

Step 1: Understand Core Concepts

What is a Prompt? Any input you provide to an AI model to generate a response:

•Question: "What is photosynthesis?"
•Instruction: "Write a haiku about coding"
•Complex request with context, examples, and constraints

Key AI Limitations:

•Hallucinations: May generate plausible but false information
•Biases: Reflects training data biases
•Knowledge cutoffs: Limited to training timeframe
•Context windows: Limited conversation memory
•Inconsistency: Same prompt may yield different results

Step 2: Master Prompt Components

1. Context - Provide background:

code

I'm a software developer working on a React application.
I need to implement user authentication.

2. Task/Instruction - State what you want:

code

Explain JWT tokens vs session-based authentication,
including pros and cons.

3. Constraints - Specify limits:

code

Keep under 200 words. Use simple language. Format as table.

4. Examples - Show patterns (see Few-Shot):

code

Input: "happy" → Output: "positive"
Input: "frustrated" → Output: "negative"

Step 3: Core Techniques

Zero-Shot Prompting Ask without examples. Best for: general knowledge, simple tasks.

code

Classify sentiment: "The product arrived late and damaged."

Few-Shot Prompting Provide 2-5 examples to establish pattern. Best for: structured extraction, formatting.

code

Extract product and price:
"iPhone 14 - $999" → {"product": "iPhone 14", "price": 999}
"Galaxy S23 for $1199" → {"product": "Galaxy S23", "price": 1199}
"Pixel 7 only $599!" → ?

See references/few-shot-prompting.md for details.

Chain-of-Thought (CoT) Guide step-by-step reasoning. Best for: math, logic, debugging.

code

Problem: Train travels 120 km in 2 hours, then 180 km in 1.5 hours.
Average speed for entire journey?

Let's solve step by step:

See references/chain-of-thought.md for details.

Role Prompting Assign expertise to influence tone and depth.

code

You are an experienced cybersecurity consultant.
Explain zero-trust architecture to a non-technical CEO.

Step 4: Advanced Techniques

Self-Criticism Generate → Critique → Revise for higher quality.

code

Task: Write a function to validate emails.

Step 1: Initial solution
Step 2: Critique for edge cases, errors
Step 3: Revised solution

Token cost: ~2-3x. Best for: critical tasks, quality refinement.

Decomposition Break complex problems into sub-problems.

code

Problem: Design scalable URL shortener.

Sub-problem 1: Core functionality?
Sub-problem 2: Generate short URLs how?
Sub-problem 3: Data storage?
Sub-problem 4: Handle high traffic?
Final: Combine solutions

Token cost: ~1.5-3x. Best for: complex systems, architecture.

Ensembling Generate multiple solutions, select most consistent.

code

Problem: [Math problem]

Path 1: [Method A] → Answer X
Path 2: [Method B] → Answer X
Path 3: [Method C] → Answer Y

Majority: X (2/3 agree)
Final: X

Token cost: ~3-5x. Best for: critical accuracy, verification.

Thought Generation Explore multiple approaches before selecting.

code

Problem: Plan 3-day Paris trip, $500 budget.

Branch A: Culture-focused → Cost $280, Pros/Cons
Branch B: Food-focused → Cost $535, Pros/Cons
Branch C: Balanced → Cost $305, Pros/Cons

Evaluate & select: Branch C (best trade-offs)

Token cost: ~3-10x. Best for: planning, strategy, optimization.

See references/advanced-techniques.md for comprehensive coverage.

Step 5: Best Practices

•Be specific: "List 5 health issues in Golden Retrievers" vs "Tell me about dogs"
•Use delimiters: Separate Context / Task / Format clearly
•Specify output format: "As JSON", "As numbered list", "As table"
•Iterate: Start simple, add constraints based on results
•Combine techniques: Few-Shot + CoT, Decomposition + Ensembling

Step 6: Avoid Pitfalls

•Don't assume perfect knowledge or real-time data
•Don't overload with too many instructions
•Don't ignore context window limits
•Always verify critical information
•Define ambiguous terms ("better" means what?)
•Test for consistency and edge cases
•Reserve advanced techniques for tasks that need them

Examples

Content Summarization

code

Role: Executive assistant
Task: Summarize article in 3 bullet points, actionable insights
Article: [text]
Format: Bullets starting with action verbs

Code with Decomposition

code

Explain this Python function using decomposition:
def fibonacci(n): ...

1. Function purpose?
2. Base cases?
3. Recursive case?
4. Time complexity?
Combine into beginner explanation.

Problem with Self-Criticism

code

Design API caching strategy.
Step 1: Propose design
Step 2: Critique (bottlenecks? edge cases? invalidation?)
Step 3: Improved design

Decision with Thought Generation

code

Decision: Migrate to microservices?
Branch 1: Assume 5 devs
Branch 2: Assume 20 devs
Branch 3: Assume tight budget
Evaluate each, recommend based on context.

Technique Selection Guide

Technique	Token Cost	Accuracy	Use When
Zero-Shot	Low	Medium	Simple tasks
Few-Shot	Medium	High	Pattern matching
Chain-of-Thought	Med-High	High	Math, logic
Self-Criticism	High (2-3x)	Very High	Quality critical
Decomposition	Med-High (1.5-3x)	High	Complex problems
Ensembling	Very High (3-5x)	Very High	Accuracy critical
Thought Generation	Very High (3-10x)	High	Planning, strategy

Strategy: Start simple → Add CoT if needed → Advanced for critical tasks.

Use Cases by Domain

Business: Emails, proposals, marketing copy, meeting summaries, strategy (thought generation)

Development: Code explanation, debugging (decomposition), test generation, documentation

Research: Paper summaries, theory comparison (ensembling), concept explanation (decomposition)

Education: Lesson plans, practice problems, explanations, feedback (self-criticism)

Personal: Trip planning (thought generation), correspondence, learning, brainstorming

Advanced Topics

Explore references directory for deep dives:

•few-shot-prompting.md - Pattern learning techniques
•chain-of-thought.md - Reasoning and decomposition
•advanced-techniques.md - Self-criticism, ensembling, thought generation, decomposition
•prompt-templates.md - Reusable templates
•model-specific-tips.md - Optimize for ChatGPT, Claude, Gemini
•evaluation-testing.md - Measure and improve performance

Quick Reference

Technique	Trigger	Example
Instruction	Direct	"Translate...", "List..."
Role	"You are..."	"You are an expert..."
Zero-Shot	Question	"What is...?"
Few-Shot	Examples	Input→Output, Input→Output, Input→?
CoT	"Step by step:"	"Let's solve step by step:"
Self-Criticism	"Generate, critique, revise"	Step 1/2/3 pattern
Decomposition	"Sub-problems"	Sub-problem 1, 2, 3...
Ensembling	"Multiple paths"	Path 1/2/3, select best
Thought Gen	"Multiple branches"	Branch A/B/C, evaluate

Troubleshooting

Vague responses → Add constraints, use role prompting, provide examples

Hallucinations → Use CoT reasoning, self-criticism, ensembling for verification

Inconsistent format → Few-shot with examples, explicit structure specification

Refusals → Rephrase to clarify legitimate use, check safety filter triggers

Too long/short → Specify exact word/sentence count

Not optimal → Use thought generation (explore alternatives), decomposition (tackle complexity)

Too complex → Decompose into sub-problems, solve independently, combine

Combining Techniques

Powerful combinations:

•Few-Shot + CoT: Examples with reasoning
•Decomposition + Ensembling: Solve sub-problems multiple ways
•Self-Criticism + Thought Generation: Critique each branch
•All Advanced (critical tasks): Decompose → Thought Gen each sub-problem → Ensemble solutions → Self-Criticize final

Warning: Advanced combos can cost 20-50x tokens. Reserve for genuinely critical tasks.

Resources

Primary source: Learn Prompting (https://learnprompting.org)

Complete citations: See metadata.yaml

Deep dives: See references/ directory (6 detailed guides)

Performance Notes

•Longer prompts = higher cost + time
•CoT: ~2x tokens
•Self-Criticism: ~2-3x tokens
•Decomposition: ~1.5-3x tokens
•Ensembling: ~3-5x tokens
•Thought Generation: ~3-10x tokens
•Advanced combos: ~20-50x tokens

Use advanced techniques strategically where quality justifies cost.