🎭 ROLE: "You are an expert [domain]..."
📝 CONTEXT: "I'm working on [project]..."
🎯 TASK: "Please [specific request]..."
⚠️ CONSTRAINTS: "Use [format]. Do not [restrictions]..."
📚 EXAMPLES: "Here are 2 examples: ..."Chain-of-Thought prompting encourages the model to reason step-by-step before giving a final answer. This dramatically improves performance on arithmetic, logic, and multi-step reasoning tasks.
Q: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?
A: 11Q: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?
A: Let's think step by step.
Roger starts with 5 balls.
2 cans of 3 balls each = 2 * 3 = 6 balls.
5 + 6 = 11.
Answer: 11| Scenario | Standard Prompt | CoT Improves? |
|---|---|---|
| Math word problems | ~30-40% | ✅ Yes (80-90%) |
| Logical reasoning | ~40-50% | ✅ Yes (70-80%) |
| Factual QA | ~70-80% | ➖ Minimal benefit |
| Creative writing | Good | ➖ May over-rationalize |
Just add "Let's think step by step" to the end of any prompt. This often improves reasoning without providing examples!
ToT generalizes CoT by exploring multiple reasoning paths simultaneously. Instead of one chain, the model explores branches, evaluates each, and selects the best path.
You are solving a complex problem. Follow this process:
1. Generate 3-5 possible next steps
2. Evaluate each step (score 1-10)
3. Select the best step
4. Repeat until solution is found
Problem: [Your complex problem here]
Think through multiple paths before deciding.Run the same prompt multiple times (with temperature > 0) and take the majority answer. This leverages the wisdom of crowds and works especially well for factual questions.
import openai
def self_consistency(prompt, n=5, temperature=0.7):
responses = []
for _ in range(n):
response = openai.ChatCompletion.create(
model="gpt-4",
messages=[{"role": "user", "content": prompt}],
temperature=temperature
)
responses.append(response.choices[0].message.content)
# Find most common answer (simplified - use majority voting)
from collections import Counter
return Counter(responses).most_common(1)[0][0]ReAct interleaves reasoning traces with actions. The model can think, then take actions (search, calculate, call APIs), then think again based on results.
You have access to these tools:
- search(query): Search the web for information
- calculate(expression): Evaluate math expressions
- finish(answer): Submit final answer
Use this format:
Thought: I need to find X
Action: search[X]
Observation: [results from search]
Thought: Based on this, I can...
Action: calculate[...]
Observation: [calculation result]
Thought: I have the answer
Action: finish[final answer]
Question: What is the current population of France divided by 2?| Aspect | CoT | ReAct |
|---|---|---|
| Tool use | ❌ No | ✅ Yes |
| Information retrieval | ❌ Limited (knowledge cutoff) | ✅ Can search web |
| Action execution | ❌ No | ✅ Yes (APIs, code) |
| Hallucination reduction | Low | High (grounded in observations) |
| Task Type | Recommended Temperature | Why? |
|---|---|---|
| Factual QA, classification | 0.0 - 0.2 | Deterministic, consistent answers |
| Code generation | 0.2 - 0.4 | Mostly accurate with slight variation |
| Creative writing, brainstorming | 0.7 - 1.0 | More diverse, creative outputs |
| Translation, summarization | 0.3 - 0.5 | Balance of accuracy and variety |
Long prompts cost more tokens. Compress while preserving meaning:
Use AI to generate and evaluate prompts. Provide the LLM with examples of desired inputs and outputs, and ask it to write the most effective prompt.
System: You are a prompt engineer. Given example inputs and outputs, write a prompt that produces this behavior.
Example input: "I'm sad today"
Example output: {"sentiment": "negative", "confidence": 0.9}
Example input: "This is amazing!"
Example output: {"sentiment": "positive", "confidence": 0.95}
Write a prompt that would generate these outputs from any input.Test both prompts on a logic puzzle and compare results:
"A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?"
Write code that asks GPT-4 the same question 5 times with temperature 0.7 and takes the majority answer. What question would you test?
import openai
from collections import Counter
def majority_vote(question, n=5):
answers = []
for i in range(n):
response = openai.ChatCompletion.create(
model="gpt-4",
messages=[{"role": "user", "content": question}],
temperature=0.7
)
answers.append(response.choices[0].message.content)
return Counter(answers).most_common(1)[0][0]Design a ReAct prompt for a customer support agent that can search a knowledge base and escalate to a human.
You are a customer support AI with these tools:
- search_kb(query): Search internal knowledge base
- escalate_ticket(issue, priority): Escalate to human agent
Thought: User asks about refund policy
Action: search_kb["refund policy"]
Observation: 30-day refund policy, restocking fee applies
Thought: I have the answer
Action: finish["Our refund policy allows returns within 30 days..."]
If confidence < 80% or issue is urgent:
Thought: This requires human review
Action: escalate_ticket["complex billing issue", priority="high"]