How can we phrase prompts in ways to get reliably good results. There are some established techniques that have empirically improved AI response as follows:

1. Zero-shot prompting: Give AI a task with no examples.

    Write a polite text message to my coach saying I can’t make training tonight.

2. K-shot prompting: Give AI a task with at least one example.

    Classify each message as: (A) urgent, (B) soon, (C) can wait.
    Example: “Project due in 2 hours” -> A
    Example: “Can you send that file by tomorrow?” -> B
    Now classify: “Whenever you get time, review my draft.”

3. Zero-shot Chain of thought prompting: Instruct AI to reason step-by-step but give no examples.

    A gym membership costs $40 per month plus a $60 sign-up fee.
    If I paid $180 total, how many months did I pay for?
    
    Think step by step before giving the final answer.

4. Multi-shot chain of thought prompting: Instruct AI to reason step-by-step with examples that include reasoning steps. In other words, you teach AI ‘how to think’ by example.

    I want you to debug code using the same reasoning style as the examples below.
    For each case:
    -Identify the likely cause of the bug
    -Explain why it happens
    -Provide the corrected code
    -Briefly state how to avoid this mistake in the future
    
    Example 1:
    Code:
    numbers = [1, 2, 3]
    print(numbers[3])
    
    Analysis:
    Cause: Index out of range error.
    Why: List indices go from 0 to 2. Index 3 does not exist.
    Fix: 
    print(numbers[2])
    Avoid: Always check list length or use len() before indexing.
    
    Now debug:
    total = 0
    for i in range(1, 5):
        total = total + i
    print(i)
    Use the same reasoning structure.

5. Tool: Instruct AI to use tool to provide a accurate and valid response.

    Look up the opening hours for the nearest public library to Parramatta NSW and tell me if it’s open this Saturday afternoon.

6. RAG (Retrieval-Augmented Generation): RAG is a method that lets AI answer questions using specific, retrieved documents instead of relying only on its internal training data.

   Retrieved Context + User prompt = RAG Prompt.

   Example of RAG Prompt:

    Answer using ONLY the information provided below.
    If the information is insufficient, say so.
    
    Context:
    Premium subscriptions include unlimited downloads and priority support.
    Refunds are available within 14 days of purchase.
    After 14 days, refunds are not permitted.
    
    Question: I bought premium 20 days ago. Can I get a refund?

7. Reflexion: Reflexion prompting is a structured self-review loop where the model improves its own answer through critique and revision.

   Example 1:

    Solve this logic problem.
    After giving your answer, list possible mistakes in your reasoning.
    Then re-solve it more carefully.

   Example 2:

    Recommend whether I should lease or buy a car.
    Then identify assumptions you made in your reasoning.
    Then provide a revised recommendation considering alternative assumptions.