Play with prompt engineering: a practical manual for making AI understand human language

1. Three Keys to Model Parameter Adjustment 

1. Temperature

• Low temperature (0.1): "The sky is __" → "Blue" (high certainty)
• High temperature (0.9): "The sky is __" → "A canvas strewn with stars" (high creativity)

2. Top-K & Top-P

Practical parameter settings:

• Exact answer tasks:temperature=0, top_k=1(Strict prediction mode)
• Creative copywriting tasks:temperature=0.9, top_k=40, top_p=0.95(Divergent thinking mode)
• General question-answering tasks:temperature=0.2, top_k=30, top_p=0.8(Balanced Mode)

2. Comparison of prompt word formats 

The same needs, different questioning methods, and different results:

# Experiment with three types of questions
prompts = [
    # Question type
    "What is the Sega Dreamcast and why is it considered a revolutionary gaming console?" ,
    
    # Declarative
    "Sega Dreamcast is the sixth generation game console released in 1999..." ,
    
    # Command type
    "Write a paragraph about the Sega Dreamcast and why it is revolutionary in the history of video game consoles."
]

# Output three different results for comparison
for  prompt  in  prompts:
    response = model.predict(prompt, temperature= 0.7 )
    print( f"Prompt type:  {prompt[: 10 ]} ...\nOutput:  {response} \n---" )

3. The Secrets of the Three Prompt Modes 

1. System Prompting

# Movie review classifier
prompt =  """
Classify movie reviews as positive, neutral, or negative.
Only uppercase tags are returned.
Review: Her is a disturbing film about where artificial intelligence is headed.
It was so painful for me to watch that I couldn't bear to watch it.
emotion:
"""

2. Role Prompting

# Travel Advisor Role
prompt =  """
I want you to play a travel consultant.
I will tell you where I am and you need to recommend me 3 places with a humorous style.
My request: "I'm in Manhattan"
"""
# Output: The Empire State Building, MoMA, and Fifth Avenue will be introduced in a humorous tone

3. Contextual Prompting

# Retro Gaming Blog
prompt =  """
Context: You're writing for a blog about 80s arcade games.
Suggest 3 article topics, each with an accompanying brief description.
"""
# Output: Evolution of arcade cabinet design, iconic 80s games, the rise and resurgence of pixel art

4. Advanced Thinking Skills 

1. Chain of Thought (CoT) Practice

Wrong example : "When Xiaoming was 3 years old, his sister was 3 times his age. Now Xiaoming is 20 years old. How old is his sister?" → Wrong answer: 63 years old

Correct posture :

Please think in steps:
1.  When Xiao Ming was 3 years old, his sister’s age was: 3 × 3 = 9 years old
2.  Age difference: 9 - 3 = 6 years old
3.  Xiao Ming is now 20 years old, and his sister is: 20 + 6 = 26 years old

2. Step-back Prompting

# Step 1: Take a step back and think about more general issues
response1 = model.predict( "Which scenes are suitable for level design of first-person shooter games?" )
# Output: Abandoned military base, cyberpunk city, alien spaceship, zombie town, underwater research facility...

# Step 2: Based on the general answer, return to the specific question
prompt2 =  f"""
Consider the following scenario:
{response1}

Now, design a compelling level story for a first-person shooter game.
"""
response2 = model.predict(prompt2)
# Output: A game level design with more depth and detail

3. Tree of Thoughts

Let’s take an example:

4. Self-consistency

# Use multiple inferences to verify the answer
responses = []
for  i  in  range( 5 ):   # Run the same problem 5 times
    response = model.predict(
        "To determine whether the content of an email is important or not, you need to think in steps:\n[email content]\nJudgement result:" ,
        temperature = 0.8 # Use high temperature to produce diverse results  
    )
    responses.append(response)

# Count the most results
import  collections
result = collections.Counter([r.split( "Conclusion:" )[ -1 ].strip()  for  r  in  responses]).most_common( 1 )[ 0 ][ 0 ]
print( f"Final judgment:  {result} " )

5. ReAct mode (reasoning + action)

from  langchain.agents  import  load_tools
from  langchain.agents  import  initialize_agent

# Initialize the search tool
tools = load_tools([ "serpapi" ], llm=llm)
agent = initialize_agent(
    tools, 
    llm,
    agent= "zero-shot-react-description" ,
    verbose= True
)

# Execute multi-step query
agent.run( "How many children do the members of the band Metallica have?" )

Execution process:

Thoughts: Metallica has 4 members
Action: Search for  "How many kids does James Hetfield have?"
Observation: three children
Thinking: 1/4 of the members have 3 children
Action: Search for  "How many kids does Lars Ulrich have?"
Observations: 3
Thinking: 2/4 members have 6 children
...Final answer: 10

5. Code Tips Expert Tips 

1. Generate code

prompt =  """
Write a Bash script that asks for a folder name and copies all the files in that folder to
Rename by adding 'draft_' prefix to the original file name.
"""

2. Debugging Code

prompt =  """
The following Python code has an error:
```Python
import os
import shutil
folder_name = input("Please enter the folder name: ")
prefix = input("Please enter a prefix: ")
text = toUpperCase(prefix) # Error here
...

Please diagnose the problem and provide the correct solution. """

## 6. Structured Output Black Technology
```json
{
  "Movie Reviews" : {
    "Sentimental orientation" :  "POSITIVE" ,
    "Reasons" : [ "Excellent acting" ,  "Compact plot" ],
    "Recommendation Index" : 4.5
  }
}

Advantages :

• Reduced hallucination content by 30%.[3]
• Force the model to answer by structure
• Easy backend processing and integration

7. A complete collection of practical reminder word templates 

# Role Playing Template
As [professional role], please [perform tasks], requesting [specific requirements].
Example: As a financial advisor, analyze the key indicators of the following quarterly report, focusing on profit growth points.

# Step decomposition template
Please follow the steps below to complete [task]:
1.  First of all...
2.  Next...
3.  Finally...

# Comparative analysis template
Please compare the similarities and differences between [A] and [B] in the following aspects:
-  Cost-effective
-  Difficulty of implementation
-  Long-term value

8. Pitfall Avoidance Guide 

1. Few samples suggest to mix types:

Good example:
Positive example: "This restaurant has great service!" → POSITIVE
Counterexample: "The food was cold" → NEGATIVE
Neutral: "The food was served at an average speed" → NEUTRAL

1. Use directives instead of restrictions :

✅ Good example: "Generate a paragraph about Sega DC, only discussing hardware specifications and release date"
❌ Bad example: "Generate a paragraph about Sega DC, don't mention games or prices"

1. Record prompt experimental results :

| Prompt Name | Target | Model | Temperature | Top-K | Top-P | Prompt Content | Output Result |
|--------|------|-----|-----|-------|------|--------|--------|
| Game review v1 | Analysis of game features | gemini-pro | 0.2 | 40 | 0.8 | ... | ...