The First Scientific Taxonomy of Prompt Engineering

1. The Chaos Before Order: Prompt Engineering as Digital Alchemy

For years, the field of prompt engineering has operated more like digital alchemy than a formal science. Practitioners have relied on a collection of anecdotal tricks, intuition, and what can only be described as “vibe prompting” to coax desired outputs from large language models.

The result, as documented in “The Prompt Report” by Schulhoff et al., has been a landscape suffering from “conflicting terminology” and “a fragmented ontological understanding.” This lack of scientific rigor is not merely academic; it has severely hampered the community’s ability to communicate and replicate results.

For instance, in the foundational GPT-3 paper by Brown et al. (2020), the input Translate English to French: llama was analyzed with “llama” being the “prompt” and “Translate English to French:” being the “task description.” More recent works, including “The Prompt Report,” refer to the entire string as the prompt. This fundamental inconsistency in terminology demonstrates the pre-scientific state of the field, one full of recipes and rituals but lacking a periodic table.

2. The Solution: The First Systematic Taxonomy

The publication of “The Prompt Report” marks a pivotal moment: a definitive solution to this chaos. The paper presents the most comprehensive survey on prompt engineering to date, establishing a structured, scientific foundation for the discipline.

Research NoteFor those who enjoy the technical details...

By meticulously reviewing the literature, the authors have assembled a foundational taxonomy of 58 LLM prompting techniques, and 40 techniques for other modalities, alongside a standardized vocabulary of 33 key terms. This work provides the common language and organizational principles that were desperately needed. It signals the transition of prompt engineering from a craft into a formal science, moving us from the ambiguous world of “Alchemy” to the structured, predictable one of “Chemistry.”

3. A Structural Guide to Prompting

At the heart of “The Prompt Report” is a taxonomy organizing 58 distinct techniques into major categories, providing a structural guide to the fundamental strategies for interacting with generative AI.

1. In-Context Learning (ICL)

In-Context Learning is the ability of a model to learn skills and tasks from instructions or examples—known as exemplars—provided directly within the prompt itself, without the need for weight updates/retraining.

Few-Shot Prompting: Providing the model with a few exemplars to demonstrate the task.
Zero-Shot Prompting: Providing zero exemplars, relying solely on a directive or instruction to guide the output.

2. Thought Generation

Thought Generation techniques are designed to prompt a model to articulate its reasoning while solving a problem. The cornerstone of this category is Chain-of-Thought (CoT) Prompting. This can be implemented in a few-shot manner (with reasoning paths in exemplars) or zero-shot (by appending “Let’s think step by step”).

3. Decomposition

Decomposition involves breaking down complex problems into a series of simpler, more manageable sub-questions.

Least-to-Most Prompting: Breaks a problem into steps and solves each sub-problem sequentially.
Plan-and-Solve Prompting: Instructs the model to first devise a multi-step plan and then execute it.

4. Ensembling

Ensembling uses multiple prompts to solve the same problem and aggregates the responses.

Self-Consistency: Generates several diverse reasoning paths for the same question. The final answer is determined by a majority vote, improving robustness on complex reasoning tasks.

4. What Actually Works: A Meta-Analysis of Performance

Beyond creating a taxonomy, the paper provides critical empirical data by benchmarking several techniques on the MMLU dataset. This analysis offers a clear, data-driven look at which strategies are most effective and which may be overhyped.

The Winners: Statistically Significant Gains

The study’s results show that performance generally improves as prompting techniques become more complex. The highest-performing technique in the benchmark was Few-Shot CoT, which achieved a 0.692 accuracy score. This finding underscores the value of providing both explicit examples (few-shot) and a structured reasoning process (CoT).

The Myths: When Simpler is Better

Crucially, the benchmark also challenges common assumptions. The study found that Zero-Shot-CoT performed significantly worse than the simpler baseline Zero-Shot prompt, achieving an accuracy of 0.547 compared to 0.627. This result establishes a critical principle: prompting techniques are not universally beneficial and must be empirically validated against simpler baselines.

5. The Prompt Engineer’s Cheatsheet

To make this new scientific standard practical, the paper’s findings can be distilled into quick-reference guides.

Table 1: The Prompting Technique Hierarchy

Category	Technique	Best Use Case
Decomposition	Skeleton-of-Thought	Accelerating answer speed through parallelization.
Decomposition	Tree-of-Thought	Tasks requiring search and planning.
Self-Criticism	Self-Refine	Iteratively improving an answer based on self-generated feedback.
Ensembling	Self-Consistency	Improving accuracy on reasoning tasks by using a majority vote over multiple reasoning paths.

Table 2: Standardizing the Vocabulary

Ambiguous/Old Term	Standardized Term/Concept	Rationale
”Context”	Additional Information	The paper discourages “context” as it is overloaded.
”Task description”	Directive	The entire input string is the Prompt; the instruction is the Directive.
”Persona Prompt”	Role	”Role” is the primary term; “persona” is conflicting and overloaded.

6. Beyond Text: The Expansion into Agents and Multimodality

The principles of structured prompting are not confined to text. “The Prompt Report” documents the expansion of these foundational concepts into more complex and varied applications.

Multimodal Prompting

Prompting is increasingly evolving beyond text-based domains to include other media, such as Image Prompting and Audio Prompting. These methods apply the core logic of prompting to new sensory inputs, opening up novel capabilities for generative models.

Agentic Prompting

An Agent is defined as a GenAI system that serves a user’s goals via actions that engage with systems outside the GenAI itself. This represents a significant leap in capability, moving from simple text generation to complex problem-solving involving memory, planning, and external tools (e.g., calculators, search engines). Techniques such as ReAct (Reasoning and Acting) and Reflexion showcase this ability.

7. The Verdict: The Dawn of a Formal Discipline

”The Prompt Report” does more than just survey a field; it provides the foundational taxonomy and standardized terminology necessary for its maturation. By categorizing and defining the vast landscape of prompting techniques, this work establishes a common ground for researchers, developers, and practitioners.

Prompt Engineering is not dead; it just became a formal engineering discipline. The future belongs to those who understand the structure of a prompt, not just the wording.