advanced-prompting-and-adversarial-testing

Based on the research from The Prompt Report (co-authored by OpenAI, Microsoft, and Google), prompt engineering is about "artificial social intelligence"—knowing how to elicit the best performance from a model through specific structural patterns.

Core Prompting Techniques

1. Few-Shot Prompting (The Highest Value Technique)

Do not describe your requirements in prose; provide 3–5 concrete examples of input/output pairs.

• Structure: Use a common format the model recognizes from training data, such as XML tags or Q: [Input] / A: [Output].
• Placement: Put examples before the final instruction.
• Why it works: It establishes a pattern for the model to follow, which is more effective than descriptive instructions for style or formatting.

2. Task Decomposition

For complex logic, prevent the model from jumping to a conclusion. Force it to map the problem space first.

• The Prompt Phrase: "Before answering, list out the sub-problems that need to be solved first."
• Workflow:
  1. Ask for sub-problems.
  2. Have the model solve each sub-problem individually.
  3. Synthesize the final answer from those components.

3. Self-Criticism (Iterative Refinement)