Prompt engineering has emerged as a critical tool to refine AI outputs, but existing techniques are fragmented and lack a cohesive structure. The research team, led by Professor Feng Zhang from Renmin University of China and collaborators from Microsoft AI, Tsinghua University, and the National University of Singapore, published their new research on 15 March 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
This paper presents a comprehensive taxonomy of prompt engineering across four aspects: ​

1. Profile and Instruction: Defines the AI’s role and task requirements (e.g., acting as a medical expert).
2. Knowledge: Integrates real-time data retrieval to combat misinformation.
3. Reasoning and Planning: Enhances problem-solving through step-by-step logic and external tools.
4. Reliability: Reduces biases and ensures stable, ethical responses.

This taxonomy categorizes prompt engineering techniques from the perspective of their underlying principles and outlines the pipeline for designing effective prompts.
This taxonomy enables developers and users to systematically optimize prompts for diverse applications, from creative content generation to high-stakes decision-making. Furthermore, this taxonomy encompasses both fundamental and advanced prompt techniques, offering detailed guidance on prompt design. With the continuous advancement of prompt engineering techniques, this taxonomy can be further extended.

The framework bridges the gap between theoretical AI capabilities and practical deployment. For instance, in healthcare, AI agents using this taxonomy can retrieve the latest medical research, simulate diagnoses, and provide patient-specific advice. In legal settings, LLMs equipped with retrieval-augmented generation (RAG) techniques accurately reference statutes, reducing risks of flawed interpretations. The study also demonstrates transformative potential in robotics, software engineering, and creative industries, where structured prompts enable AI to autonomously execute tasks with human-like precision.

The team outlines six key research opportunities, including defenses against adversarial “prompt attacks” and domain-specific frameworks for fields like finance and education. The proposed future research directions are forward-looking and provide valuable guidance for subsequent research.
DOI:10.1007/s11704-025-50058-z