System Prompting - Types and Importance

BackStory

So, Let’s start with the analogy of - What really is System Prompting

• Analogy: Imagine you hire a chef. Before they start cooking, you tell them:

  • “I want Italian food only.”

  • “No spicy ingredients.”

  • “Explain each step while cooking.”

That’s exactly what system prompting does - it guides how the AI should behave.

Why is it important

• Okay, so we understood the analogy of system prompting . Now, lets see this diagram.

• 

  We have all heard about GIGO principle which states - Garbage In, Garbage Out .

• This is applicable not only in the real life , but also in the programming world , and so in the Prompting World .

• If you will give the bad prompting/ examples , how can you expect that the system will work well ?

  Hence ,

System prompting is like giving AI a personality, style, or rulebook before asking it questions. Once you set it, every answer it gives will follow that guide.

Styles of Prompting

1. Plain or Direct Prompting- The simplest form: you just ask the model a question directly.

   Example:

    Explain love in simple words.
    Error 404 lol
   

   Works, but the AI doesn’t have a “role” or strict structure, answers can vary in style, length, or tone.

2. Instruction-based Prompting (INST / Alpaca style)

   • Format: Instruction -> Input -> Response

   • Originated with Alpaca, which adapted ideas from LLaMA and FLAN.

   • The AI is told exactly what to do, then given the input, then it writes the output.

   • Structure:

       ### Instruction:
       Explain the concept of photosynthesis.
     
       ### Input:
       N/A (or any extra context)
     
       ### Response:
     

   • Use: Best for fine-tuning models to follow clear instructions.

   • You can read their article - https://crfm.stanford.edu/2023/03/13/alpaca.html

3. FLAN-style prompting

   • Used in FLAN-T5 and similar models.(Google)

   • Focused on instruction-following across a wide variety of tasks.

   • Often simpler than Alpaca, but same idea: give explicit instructions instead of vague questions.

   • Works great for zero-shot tasks (asking the model to do things without prior examples).

4. ChatML (Chat Markup Language)

• Designed for chat-based models like GPT.

• Adds roles like system, user, assistant.

• Widely adopted by all .

• Example:

    [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Explain black holes simply."},
      {"role": "assistant", "content": ""}
      {"role": "developers", "content": ""}
      {"role": "assistant", "content": ""}
    ]
  

• Why it exists: Helps the AI distinguish who is speaking, keeps conversation context clear, allows multi-turn dialogue naturally.

Types of Prompting

1. Zero-shot prompting

• Idea: The model gets no examples - just instructions. It has to answer based on what it already knows.

• Example:

    Explain why the sky is blue.
  

• Use: Quick questions, tasks where you trust the model’s knowledge.

• Analogy: You ask a stranger for directions without showing them a map.

2. Few-shot prompting

• Idea: You give the model a few examples before asking the real question. This helps it understand the pattern.

• Example:

    Q: Capital of France?  
    A: Paris  
    Q: Capital of Germany?  
    A: Berlin  
    Q: Capital of Italy?  
    A: Nahii pata bhaiiiii .lol
  

• Use: When you want the model to mimic a specific pattern or style.

• Analogy: You show the stranger 2-3 maps first, so they understand how to read yours.

3. Chain-of-Thought (CoT) prompting

• Idea: Ask the model to think step by step before answering. This improves reasoning.

• (What you see the recent feature in CHATGPT/ DEEPSEEK - Think Longer) - It uses the same concept

• Example:

    Solve: 23 + 47  
    Explain step by step.
  

• Model might respond:

    Step 1: Add 20 + 40 = 60  
    Step 2: Add 3 + 7 = 10  
    Step 3: 60 + 10 = 70
  

• Use: Math, logic, multi-step reasoning problems.

• Analogy: You ask a student to show their work, not just give the answer.

4. Self-Consistency prompting

• Idea: Ask the model to generate multiple reasoning paths and pick the most consistent answer.

• Example: For a tricky question:

    What is the next number in the series: 2, 4, 8, 16, ?
  

  The model might generate multiple chains:

  • Chain 1: Multiply by 2 → 32

  • Chain 2: Add 2 → 18

  • Chain 3: Square previous → 256

Then it selects the answer that appears most consistent across chains → 32.

• Use: Reduces mistakes, increases accuracy in complex tasks.

• Analogy: Ask several experts independently and go with the majority answer.

So, in the end , I want to say that -

“Prompting isn’t about telling AI what to say . it’s about teaching it how to think, one question at a time.”