The Prompt Report
Citation
Schulhoff, S., Ilie, M., Balepur, N., et al. (2024). The Prompt Report: A Systematic Survey of Prompting Techniques. arXiv:2406.06608.
Core Contribution
The most comprehensive systematic review of prompting techniques to date. Analyzed 1,565 papers using PRISMA methodology to create a taxonomy of 58 text-based prompting techniques, 40 multimodal techniques, and a vocabulary of 33 standardized terms. Led by researchers from University of Maryland, OpenAI, Stanford, and Microsoft.
Prompting Technique Taxonomy
Six Major Categories:
- Zero-Shot: No examples provided (role prompting, emotion prompting, rephrase and respond)
- Few-Shot: Examples included in prompt (demonstration selection, chain-of-thought)
- Thought Generation: Encourage reasoning steps (CoT, Tree of Thoughts, Graph of Thoughts)
- Ensembling: Multiple prompts/outputs combined (self-consistency, DENSE)
- Self-Criticism: Model critiques own output (self-refine, self-verification)
- Decomposition: Break complex tasks into subtasks (least-to-most, plan-and-solve)
Key Findings
- Chain-of-thought prompting significantly enhances math/reasoning tasks
- Zero-shot CoT (“Let’s think step by step”) surprisingly effective
- Prompt engineering suffers from conflicting terminology
- Few-shot example selection matters as much as technique choice
Strategic Value for heyMHK
Provides the canonical reference for discussing prompting systematically. Essential for Engine Room articles on prompting and for evaluating AI interaction patterns.
Related: 05-molecule—chain-of-thought-prompting, 05-atom—few-shot-cot-superiority, 07-source—engine-room-series