1. Introduction and Definition

Prompt engineering represents the systematic process of structuring instructions, context, constraints, and examples to guide a generative model's response toward a defined objective. This process encompasses "interactive skills" (defining role, goal, audience, output format) through to "design patterns" and "evaluation protocols". Recent systematic reviews demonstrate that prompt engineering now possesses a standardized set of techniques and is no longer merely intuitive trial-and-error.

2. Theoretical Foundations and Evolution

Since the emergence of large language models (LLMs), two key ideas have driven qualitative leaps: (a) in-context few-shot learning for implicit task and style transfer, and (b) inducing Chain-of-Thought (CoT) reasoning to compel models to articulate intermediate reasoning steps. The seminal work by Wei and colleagues demonstrated that with just a few CoT examples, reasoning capability in computational and logical problems dramatically improves—particularly in larger models.

Concurrently, the literature on "prompt patterns" emerged, inspired by software design patterns, cataloging reusable solutions (such as "role assignment," "explicit format constraints," "scaffolded examples," "critical questioning") to enhance knowledge transferability.

3. Classification of Methods and Patterns

3.1 Foundational Methods

Zero-shot with clear instructions and sufficient context
Few-shot for inducing style/pattern
Role prompting for controlling perspective and output norms
Format constraints (Schema/Format) for structured JSON/table/report generation

Reviews have shown these methods form the core of most general workflows.

3.2 Reasoning Methods

Chain-of-Thought (CoT): Inducing step-by-step reasoning
Self-consistency: Multiple executions with voting for consistent answer selection
Program-of-Thought/Procedural structuring: Breaking problems into functions/steps and compelling model adherence

Literature demonstrates CoT is particularly effective in multi-step and computational problems, while self-consistency reduces error rates.

3.3 Data-Driven and Knowledge-Driven Methods

Retrieval-Augmented Generation (RAG): Injecting up-to-date evidence into prompt context to reduce hallucinations
Pre-/Meta-Prompting at organizational level: Defining governance/header prompts that enforce norms, authorized sources, and legal constraints (e.g., corporate settings)

Strategic documents in Germany recommend these approaches for compliance and traceability.

3.4 Design Patterns (Prompt Patterns)

Pattern catalogs include "controlled examples," "task decomposition," "constrained rewriting," "Socratic questioning," and "self-evaluation with criteria," enabling selectivity and repeatability.

4. Metrics and Evaluation

4.1 Intrinsic Output Metrics

Factuality: Verification against external sources/human judgment
Coverage/Adequacy relative to task criteria
Format compliance (e.g., valid JSON, length/style)

Recent reviews emphasize the necessity of shared vocabulary for these metrics.

4.2 Process Metrics

Robustness against minor text variations
Cost and time (token count/interaction rounds)
Auditability: Ability to reproduce results with logged prompt and model version

4.3 Evaluation Designs

A/B testing prompts with matched samples
Self-consistency with voting
Human-in-the-loop evaluation for qualitative tasks

Comprehensive reviews recommend hybrid evaluation (automated + human).

5. Risks and Threats

5.1 Prompt Injection and Jailbreaking

Any input text (even links/files) can carry adversarial instructions and circumvent policies. Mitigation strategies include separating data/instruction channels, ignoring unauthorized text, and secure prompt rewriting. German guidelines for organizational environments also emphasize input policies, access control, and documentation.

5.2 Bias, Factual Errors, and Over-Reliance

Mitigation through RAG, mandatory citation, and human review. German academic references in higher education also emphasize "prompt literacy" and "critical review of outputs."

5.3 Emerging Abuses

Recent examples of "hiding instructional messages for LLM-based reviewers" demonstrate that prompting can manipulate scientific processes, highlighting the necessity of input text auditing and ethical review systems.

6. Governance, Ethics, and Compliance (Focus on Germany/EU)

For organizational deployment of prompt engineering, three governance layers are proposed:

Strategy and Policy: Defining scope of use, authorized sources, risk mapping based on application
Technical Controls: Header-prompt patterns, blacklist/whitelist of sources, metadata logging (prompt, model version, timestamp)
Audit and Training: "Prompt literacy" training, periodic review

In German-speaking contexts, industry (Bitkom) and academic guidelines recommend step-by-step implementation of compliance with the EU AI Act (KI-VO) and attention to data protection; these documents also reference the role of RAG and organizational pre-prompts in risk control.

7. Best Practices for Prompt Design

Make goal and audience explicit (what you want, for whom, with what tone/format)
Provide minimal sufficient context (definitions, constraints, brief example)
Lock output format (checklist, JSON, headings)
Stage the process (incremental solving, questioning, self-checking)
Use patterns (role, scaffolded examples, Socratic questioning, self-consistency testing)
Systematic evaluation (A/B, quality metrics, human review)
Safety and compliance (input filtering, mandatory citation, audit logging)

Academic/educational guidance in Germany also confirms these principles as practical recommendations for teachers and students.

8. Example Research Prompt Template (Practical Summary)

Role/Audience: "As a reviewer for Journal X…"

Task: "Evaluate the paper on criteria A/B/C…"

Context: "Paper abstract, journal criteria…"

Output Constraints: "3-part report with quantitative score and recommendation…"

Process: "First strengths, then weaknesses, then suggestions; finally perform a consistency check."

This template aligns with design patterns and findings from CoT and self-consistency research.

Conclusion

Prompt engineering today is a discipline, not merely "trial-and-error art." Research literature demonstrates that combining foundational methods (zero/few-shot, role, format constraints) with reasoning methods (CoT, self-consistency) and knowledge injection (RAG) can significantly enhance output efficiency, accuracy, and reproducibility. For organizational maturity, data and prompt governance, transparent policies, continuous evaluation, and "prompt literacy" training are essential—particularly in the German-speaking ecosystem where specific guidance infrastructure and regulations (KI-VO/EU AI Act and industry/academic guidelines) are being shaped and updated. Future research should focus on standardizing metrics, audit protocols, and integrating prompt engineering with quality assurance tools (source validation, decision traceability).

References

Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903.
White, J., et al. (2023). A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT. arXiv:2302.11382.
Chen, B., Zhang, Z., Langrené, N., Zhu, S. (2023). Unleashing the Potential of Prompt Engineering in Large Language Models: A Comprehensive Review. arXiv:2310.14735.
Schulhoff, S., et al. (2024). A Systematic Survey of Prompt Engineering Techniques. arXiv:2406.06608.
Sahoo, P., et al. (2024). A Systematic Survey of Prompt Engineering in Large Language Models. arXiv:2402.07927.
Bitkom (2024). Umsetzungsleitfaden zur KI-Verordnung (EU) 2024/1689. Berlin: Bitkom.
Bitkom (2025). Künstliche Intelligenz & Datenschutz – Praxisleitfaden (2. Auflage). Berlin: Bitkom.
Hochschulforum Digitalisierung (2024). Blickpunkt: Leitlinien zum Umgang mit generativer KI. Berlin: HFD.
TH Köln (2024). Wie Sie richtig prompten – Promptanleitung (GPT-Lab). Köln: TH Köln.
TU Darmstadt (2025). Handreichung generative KI für Studium und Lehre. Darmstadt: HDA.
Hochschulforum Digitalisierung (2023–2025). Prompt-Labor / Selbstlernmaterialien. Berlin: HFD.
Universität Osnabrück (2025). Handlungsempfehlungen zum Umgang mit KI-basierten Anwendungen. Osnabrück.
The Guardian (2025). Scientists reportedly hiding AI text prompts in academic papers. London: The Guardian.

Blog & Insights

Prompt Engineering: How to Steer AI Models Effectively—and Responsibly—with Better Prompts