Prompt Engineering Notes
Prompt engineering is the process of designing high-quality prompts to guide LLMs to produce accurate outputs. This involves refining length, structure, tone, and clarity.
Model-Specific Setup
Choose Your LLM and Configure:
- Model-specific prompts and capabilities
- Sampling parameters:
- Output Length
More tokens = more compute cost. Reducing length just truncates output. - ReAct model warning: Emits irrelevant tokens post-response.
- Output Length
Sampling Controls
| Parameter | Effect |
|---|---|
Temperature | Low = deterministic; High = creative/random |
Top-K | Limit prediction to top K likely tokens |
Top-P | Nucleus sampling = choose from top cumulative probability P |
Num Tokens | Max output length |
Practical Guidelines:
- Temperature = 0.2, Top-P = 0.95, Top-K = 30 → Balanced
- Temperature = 0.9, Top-P = 0.99, Top-K = 40 → Creative
- Temperature = 0.1, Top-P = 0.9, Top-K = 20 → Factual
- Temperature = 0 → Tasks with one correct answer (math, facts)
⚠️ Beware of infinite loops & repetition at low temperature + long max tokens.
Prompting Techniques
- Zero-Shot
- One-Shot
- Few-Shot → Include edge cases!
Role, Context, and System Prompting
| Type | Use |
|---|---|
System | Model’s base behavior, safety (e.g., “Be respectful…”) |
Role | Output style: Formal, Humorous, Persuasive, etc. |
Context | Immediate task input (e.g., “You’re writing a blog…”) |
Advanced Prompting Methods
Step-Back Prompting
- Ask a general question before the main task to activate relevant knowledge.
- Helps mitigate bias and improve performance.
Chain-of-Thought (CoT)
- Use: “Let’s think step-by-step.”
- Great for reasoning-heavy tasks.
- CoT + few-shot = even better.
Self-Consistency Prompting
- Generate multiple CoTs using high temperature.
- Extract final answers → choose the majority.
Tree-of-Thoughts (ToT)
- Explores multiple CoTs in parallel.
- Great for planning and complex reasoning.
→ Read paper
ReAct (Reason + Act)
- Combine reasoning with action (e.g., API calls, tools).
- Resend state, prompts, and reasoning continually.
Auto-Prompting & Code Prompting
- Write prompts that write better prompts.
- Use BLEU / ROUGE to evaluate prompt performance.
- Code prompting = Specify interfaces, structure, error handling, etc.
Best Practices
- Give Examples
- Be Clear & Concise
- Prefer Positive Instructions over negative constraints.
- Use strong action verbs: Act, Analyze, Generate, Rank, etc.
- Control Max Token Length
- Tune CoT setups:
- Use greedy decoding (temp = 0)
- Extract answers separately from reasoning
- Log & Document all prompt versions and results
📎 Reference: Google Whitepaper
Title: Prompt Engineering: Best Practices and Reasoning Strategies
Author: Lee Boonstra, Google Cloud
Link: Read PDF