Prompt Engineering Practical Guide: A Systematic Approach from Basics to Advanced Techniques

Prompt engineering is the practice of designing inputs to large language models to produce expected outputs. As frontier models like GPT-4 and Claude 3 series have advanced, prompt engineering has evolved from “dark arts” into a system with traceable principles. Anthropic, OpenAI, and others have published detailed prompt engineering guides.

## Foundational Principles: Clear, Specific, Structured

**Define a role**: open with a model role — “You are a clinical trial design expert with ten years of experience” — helping the model activate relevant knowledge domains and calibrate output style and professional depth.

**Specific task description**: avoid vague instructions (“write me a report”); use concrete specifications (“write an 800-word competitive analysis report with a feature comparison table, strengths/weaknesses analysis, and market positioning recommendations, targeted at non-technical management readers”).

**Structured format requirements**: explicitly specify output format — JSON, Markdown tables, numbered lists, specific XML tags — reducing arbitrary model formatting choices. Anthropic recommends XML tags for organizing complex prompt structures.

## Advanced Techniques

**Chain-of-Thought (CoT)**: adding “think step by step” or “let’s analyze this systematically” at the prompt end significantly improves model performance on reasoning, math, and logic tasks. This is especially effective with reasoning models (o1, Claude 3.7 Sonnet extended thinking).

**Few-shot prompting**: provide 2–5 input-output examples, letting the model learn the desired output pattern. Highly effective for tasks with custom formatting requirements (structured data extraction, for instance).

**Negative examples**: telling the model what NOT to do — “don’t use marketing language,” “don’t use bullet points” — is sometimes more effective than positive instructions.

**Tree of Thoughts**: have the model generate multiple solution drafts, evaluate each, and select the optimal path to continue. Suited for complex decisions requiring exploratory thinking.

## Claude-Specific Optimization

Anthropic’s prompt engineering documentation (docs.anthropic.com) covers Claude-specific optimizations: using `` tags to trigger extended thinking; XML tags like ``, `` to clearly separate content; explicitly stating the task’s purpose and audience so Claude calibrates output style. For long-document processing, placing documents before instructions (rather than after) improves citation accuracy.

See [RAG Knowledge Retrieval](https://sunqi.org/rag-knowledge-retrieval-en/) and [Anthropic prompt engineering overview](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview).