Prompt Engineering

Last reviewed: September 2026

Note

New to LLMs/prompts? Read Getting Started with AI first.

What Is Prompt Engineering

Prompt engineering addresses “how to question/instruct a model to get the desired output.” The same model produces vastly different quality depending on the prompt.

Per Anthropic’s official guide, prompt engineering is an “empirical science” — the most important thing is defining evaluation criteria → iterative testing. It’s not about finding a “good prompt” but improving through measurable metrics — Anthropic Prompt Engineering Guide

Core Principles

Common principles emphasized by Microsoft, Google, and Anthropic guides:

• Clear, specific instructions — Vague requests produce vague answers.
• Provide context — Give background information the model doesn’t have.
• Assign a role — Persona setting like “You are a legal expert.”
• Specify output format — JSON, list, paragraph, etc.
• Provide examples — Input/output examples improve quality (Few-shot).
• Iterate with evaluation — Measure quality on representative cases, then refine.

Sources:

• Microsoft — Prompt engineering techniques
• Google Cloud — Prompting strategies
• Anthropic — Prompt engineering overview

Key Patterns

Zero-shot

Direct instruction without examples. Suitable for simple tasks but quality drops on complex ones.

Analyze the sentiment of the following sentence: "Today's meeting was boring."

Few-shot (One-shot, Multi-shot)

Show desired patterns via examples first. The LLM mimics the format.

Classify each sentence as positive/neutral/negative.

Sentence: "The service was excellent."
Answer: Positive

Sentence: "Delivery arrived on schedule."
Answer: Neutral

Sentence: "The product arrived damaged."
Answer:

Sources:

• Google Cloud — Give examples (few-shot prompting)
• Anthropic — Use examples (multishot prompting)

Chain-of-Thought (CoT)

Guide the model to reason step by step. Dramatically improves accuracy on math, logic, and complex classification.

Solve this step by step.

Problem: A store has 23 apples. 20 were sold. Then 6 new ones arrived. How many are there now?

Steps:
1.

Latest frontier models (Claude, GPT, Gemini families, etc.) sometimes perform CoT internally, but explicit “Think step by step” instructions remain effective.

Sources:

• Microsoft — Chain of thought prompting
• Anthropic — Let Claude think (CoT)
• Wei et al., 2022 — Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Role Assignment (System Prompt / Persona)

Assign a role to shape tone, expertise, and response scope. Most APIs support a separate system prompt.

You are a Korean legal expert. Explain in terms a layperson can understand,
while using legal terminology accurately.

Question: What do the contract terms '갑' (gap, "Party A") and '을' (eul, "Party B") mean?

See the Anthropic guide for more on system prompts.

Structured Output

Specify desired format (JSON, XML) explicitly.

Extract information from the review below as JSON.

Schema:
{
  "product": "product name",
  "rating": "integer 1-5",
  "sentiment": "positive" | "neutral" | "negative"
}

Review: "These earbuds are amazing! Great sound quality and long battery life. Five stars."

Some vendors natively support Structured Outputs:

• Microsoft Foundry Structured Outputs
• Vertex AI Controlled Generation

ReAct (Reasoning + Acting)

The LLM alternates “think → tool call → observe → next thought” in a loop. Foundational pattern for agent implementations.

Available tools:
- search(query): web search
- calculate(expression): calculation

Question: What is the distance between the 2024 Olympics host city and Seoul?

Thought: First, I need to find the 2024 Olympics host city.
Action: search("2024 Olympics host city")
Observation: Paris
Thought: I need to calculate the distance between Paris and Seoul.
Action: search("distance Paris Seoul")
Observation: approximately 8,957 km
Answer: The 2024 Olympics host city is Paris, approximately 8,957 km from Seoul.

Sources:

• Yao et al., 2022 — ReAct
• AWS — Tool Use with Bedrock
• Anthropic — Tool use

Choosing the Right Technique

Knowing the patterns matters less than judging which technique to use when. Consider the situation, model characteristics, and cost together.

Situation	Recommended technique	Why / caution
Simple classification/extraction/transformation	Zero-shot	Sufficient without examples; minimizes token waste
Tricky output format or tone	Few-shot	“Show” the format with 2–5 examples; avoid biased examples
Multi-step reasoning, math, logic	Chain-of-Thought	Elicits intermediate reasoning. But reasoning-specialized models (o-series, reasoning modes) build in CoT, so explicit CoT is unnecessary or even counterproductive
Structured output (JSON, etc.) required	Output-format enforcement + schema	Native structured-output / function-calling APIs (where supported) are more reliable than prompt instructions
Combined with external tools/search	ReAct / tool calling	Agentic workflow; needs loop-cost and infinite-loop safeguards
Consistent role/policy	System Prompt / Persona	Fixed across the session; must not be overridden by user input

By model type:

• General models (GPT-6, Claude Sonnet, etc.): Few-shot and CoT meaningfully improve quality.
• Reasoning-specialized models (o-series, Claude reasoning mode, etc.): They reason internally, so “think step by step” instructions are unnecessary, and excessive CoT can hurt performance and cost.
• Cost angle: Few-shot examples and CoT increase input tokens. On high-frequency paths, minimize example count / reasoning length or use prompt caching.

Note

Which technique is actually better must be confirmed by evaluation. A/B-measure accuracy, format-adherence, and token cost on a representative input set, then continue with the improvement loop below.

Vendor-Specific Best Practices

Vendor	Key Recommendation	Reference
AWS Bedrock (Claude)	Structure with XML tags (<context>...</context>), clear instructions first	Claude best practices
Microsoft Foundry (GPT)	Fix role via system prompt, provide format examples	Azure prompt engineering guide
Google Cloud (Gemini)	Clear instructions, explicit constraints, iterative experimentation	Vertex AI Prompt strategies
OCI (Cohere)	Preamble for persona, precise JSON schema for tool use	Cohere Prompt Engineering

Prompt Improvement Loop

graph LR
    A[1. Define Success Criteria] --> B[2. Build Eval Set]
    B --> C[3. Draft Prompt]
    C --> D[4. Run Evaluation]
    D --> E{Meets Criteria?}
    E -->|No| F[5. Refine Prompt]
    F --> D
    E -->|Yes| G[Deploy]

A “good prompt” is never completed in one shot. Build an eval set of 20–50 representative questions and compare evaluation scores on each prompt change.

Common Mistakes

• Too many instructions in one prompt — 5+ directives are easily missed by models.
• Demanding complex formats without examples — Few-shot examples are far more effective.
• Testing only representative questions — Quality can collapse on edge cases.
• Not re-evaluating prompts on model upgrades — Same prompt may produce different output across model versions.

References

AWS

• Bedrock Prompt Engineering Guidelines

Azure

• Microsoft Foundry — Prompt engineering techniques
• Microsoft — Generative AI for Beginners

Google Cloud

• Vertex AI — Prompting strategies

Standards & Community

• Anthropic — Prompt engineering overview
• Cohere — Prompt engineering
• Wei et al., 2022 — Chain-of-Thought Prompting
• Yao et al., 2022 — ReAct
• Brown et al., 2020 — GPT-3 (Few-shot learning)
• Anthropic — Interactive Prompt Engineering Tutorial