====== AI Prompting Techniques ======

An AI prompting technique is a structured method of communicating with large language models to maximize the quality, accuracy, and usefulness of their outputs. The quality of a prompt is directly related to the quality of the response received. ((source [[https://www.k2view.com/blog/prompt-engineering-techniques/|K2view - Prompt Engineering Techniques]])) Mastering these techniques is the fastest and cheapest lever available to improve LLM performance without any fine-tuning. ((source [[https://www.datacamp.com/blog/what-is-prompt-engineering-the-future-of-ai-communication|DataCamp - What Is Prompt Engineering]]))

===== Zero-Shot Prompting =====

Zero-shot prompting instructs the model to perform a task without providing any examples, relying entirely on its pre-trained knowledge to generalize. This is the simplest prompting method.

**Best for:** Translation, summarization, sentiment analysis, content moderation, and quick prototyping. ((source [[https://www.k2view.com/blog/prompt-engineering-techniques/|K2view - Prompt Engineering Techniques]]))

**Example:** "Classify this text as neutral, negative, or positive: The product broke after just 2 days of use."

===== Few-Shot Prompting =====

Few-shot prompting provides a small number of examples in the prompt to demonstrate the desired task, enabling in-context learning for complex scenarios where zero-shot fails. Think of it as teaching a friend how to play a game by showing them a couple of moves first. ((source [[https://www.datacamp.com/blog/what-is-prompt-engineering-the-future-of-ai-communication|DataCamp - What Is Prompt Engineering]]))

**Best for:** Tasks requiring nuance or context, text classification, format consistency, and pattern-based generation.

**Example:**
  Positive -> Optimistic
  Negative -> Pessimistic
  Confident -> ?

===== Chain-of-Thought Prompting =====

Chain-of-thought (CoT) prompting encourages step-by-step reasoning by adding phrases like "Think step by step," breaking complex problems into sub-steps for better accuracy. The core insight is that when a model writes out an intermediate reasoning step, that step becomes available as context for subsequent generation. ((source [[https://mbrenndoerfer.com/writing/chain-of-thought-prompting-zero-shot-fine-tuning-limitations|Brenndoerfer - Chain-of-Thought Prompting]]))

CoT improves reasoning accuracy by 10 to 40 percent on complex tasks. ((source [[https://glyphsignal.com/guides/prompt-engineering-guide|GlyphSignal - Prompt Engineering Guide 2026]]))

**Best for:** Math problems, logic, troubleshooting, decision-making, and multi-step analysis.

**Variants:**
  * **Zero-shot CoT:** Simply add "Think step by step" without examples
  * **Few-shot CoT:** Provide worked examples showing the reasoning chain

===== Tree-of-Thought Prompting =====

Tree-of-thought (ToT) prompting extends chain-of-thought by exploring multiple reasoning paths simultaneously in a tree-like structure. The model evaluates and prunes branches to find optimal solutions, making it particularly effective for problems where multiple valid approaches exist. ((source [[https://www.lakera.ai/blog/prompt-engineering-guide|Lakera - Prompt Engineering Guide]]))

**Best for:** Complex planning problems, strategic decisions, creative problem-solving with multiple valid paths.

===== ReAct Prompting =====

ReAct (Reason plus Act) interleaves reasoning traces with actions such as tool calls or external queries. The model thinks about what it needs, performs an action to gather information, observes the result, and continues reasoning. ((source [[https://www.promptingguide.ai/techniques|Prompting Guide - Techniques]]))

**Best for:** Research tasks, fact-checking, multi-step workflows requiring external data, and agentic AI applications.

===== Self-Consistency Prompting =====

Self-consistency generates multiple reasoning paths for the same prompt and selects the most consistent answer through majority vote, improving reliability on ambiguous tasks. ((source [[https://www.k2view.com/blog/prompt-engineering-techniques/|K2view - Prompt Engineering Techniques]]))

**Best for:** Ambiguous questions, mathematical reasoning, and any task where reliability matters more than speed.

===== Role Prompting =====

Role prompting assigns the model a specific persona to align its tone, expertise, and behavior with the desired output. ((source [[https://www.lakera.ai/blog/prompt-engineering-guide|Lakera - Prompt Engineering Guide]]))

**Example:** "You are a legal advisor specializing in intellectual property law. Review this contract clause and identify potential issues."

**Best for:** Customer support, domain-specific advice, data analysis, and any task requiring a particular perspective or expertise level.

===== Meta-Prompting =====

Meta-prompting structures responses abstractly by outlining steps or formats rather than providing full examples. It focuses on the logic of how to approach a problem rather than showing completed solutions. ((source [[https://www.k2view.com/blog/prompt-engineering-techniques/|K2view - Prompt Engineering Techniques]]))

**Example:** "Step 1: Define the variables. Step 2: Apply the formula. Step 3: Verify the result."

**Best for:** Structured problem-solving, mathematical workflows, and consistent output formatting.

===== Prompt Chaining =====

Prompt chaining sequences multiple prompts across interactions, building context incrementally rather than cramming everything into a single input. Each prompt builds on the output of the previous one. ((source [[https://www.lakera.ai/blog/prompt-engineering-guide|Lakera - Prompt Engineering Guide]]))

**Best for:** Long research workflows, complex multi-phase tasks, onboarding processes, and any task too large for a single prompt.

===== Emerging Techniques (2025-2026) =====

**Multi-Turn Memory Prompting** builds layered context over sessions, effectively training the model memory for personalized, ongoing interactions. ((source [[https://www.lakera.ai/blog/prompt-engineering-guide|Lakera - Prompt Engineering Guide]]))

**Adaptive Prompting** allows models to dynamically adjust responses to the user style. Concise inputs receive concise outputs, enhancing naturalness in chatbots and assistants. ((source [[https://www.datacamp.com/blog/what-is-prompt-engineering-the-future-of-ai-communication|DataCamp - What Is Prompt Engineering]]))

**Combining Techniques** blends multiple approaches such as CoT combined with role prompting and few-shot examples for hybrid tasks requiring multifaceted guidance.

===== Choosing the Right Technique =====

^ Scenario ^ Recommended Technique ^
| Simple factual question | Zero-shot |
| Consistent output format needed | Few-shot |
| Complex reasoning or math | Chain-of-thought |
| Multiple valid solution paths | Tree-of-thought |
| Tasks requiring external data | ReAct |
| High-reliability answers | Self-consistency |
| Domain-specific expertise | Role prompting |
| Large multi-step workflows | Prompt chaining |

===== See Also =====

  * [[master_ai_prompting|How to Master AI Prompting]]
  * [[ai_prompt_guardrails|AI Prompt Guardrails]]
  * [[agentic_ai_vs_generative_ai|Agentic AI vs Generative AI]]
  * [[rag_in_ai|What Is RAG in AI]]

===== References =====