Basic of Prompt Engineering

Overview

#

Prompt engineering is the art of crafting effective instructions for Large Language Models (LLMs) to achieve desired outputs.

Core Techniques

#

1. Chain of Thought (CoT)

#

Explicit CoT: Provide step-by-step reasoning guidance.

Solve this problem step by step:
Q: If a train travels 120 km in 2 hours, what is its speed?

Think through:
1. Identify what we know
2. Apply the formula
3. Calculate the answer

Zero-Shot CoT: Let the model reason independently.

Q: If a train travels 120 km in 2 hours, what is its speed?
Let's think step by step.

2. Self-Consistency

#

Sample multiple outputs (~20) and vote on the most common answer.

Parameters:

• Temperature: Controls randomness (0.7-1.0 for diversity)
• Top-K: Limits token selection pool

Generate 20 solutions with temperature=0.8
→ Select most frequent answer

3. Sampling-and-Voting (Ensemble)

#

Use multiple models or personas:

As a mathematician, solve: ...
As a physicist, solve: ...
As an engineer, solve: ...
→ Combine answers

Smaller ensembles can outperform single large models.

4. ReAct (Reasoning + Action)

#

Interleave reasoning with actions:

Thought: I need to find the current weather
Action: search("weather today Seoul")
Observation: 15°C, cloudy
Thought: Now I can answer the user
Response: It's 15°C and cloudy in Seoul today.

5. Self-Evaluation

#

Self-Critique:

[Generate response]
Now critique your answer:
- Is it accurate?
- Is anything missing?
- How can it be improved?
[Revise based on critique]

Constitutional AI:

Evaluate if your response:
- Is helpful
- Is harmless
- Is honest

Advanced Strategies

#

Output Formatting

#

Structure responses effectively:

Best Practices

#

1. Be specific - Clear, unambiguous instructions
2. Provide examples - Few-shot learning
3. Set constraints - Length, format, style
4. Iterate - Refine prompts based on outputs
5. Use delimiters - Separate sections clearly

### Task ###
[Your task description]

### Context ###
[Relevant background]

### Format ###
[Expected output format]