Prompt engineering

Working Definition

Prompt engineering is the systematic practice of designing, structuring, and optimizing the input text (and multimodal context) fed to a generative model so as to elicit desired behaviors—such as a specific output schema, persona, reasoning trajectory, or epistemic constraint—without modifying the model’s underlying weights. In its rigorous technical sense, it involves constructing parameterized templates, curating few-shot demonstrations, orchestrating intermediate reasoning steps, and evaluating prompt variants quantitatively across held-out test suites. In popular and commercial usage since 2022, the phrase simply means “knowing how to talk to ChatGPT” and has been marketed as a standalone white-collar career or a collection of copy-paste “magic phrases.” Keeping the two apart is essential: engineering a prompt pipeline is an empirical discipline measured by variance and error rates across distributions of inputs, whereas folk prompting relies on single-shot anecdotes that frequently collapse when the model checkpoint updates.

Genesis and Global Lineage

The practice preceded the name. When OpenAI published the GPT-3 paper (Language Models are Few-Shot Learners, Brown et al., May 2020) and opened its commercial API that summer, early users and researchers discovered that rewording a natural-language task description or re-ordering a handful of input–output examples could swing benchmark accuracy by thirty or forty percentage points without touching a single gradient. In February 2021, Laria Reynolds and Kyle McDonell published Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm, coining “prompt programming” and arguing that prompts do not teach a model new concepts on the fly, but rather use natural-language cues to locate and activate task spaces the model has already absorbed during pretraining. In July 2021, Pengfei Liu and colleagues systematized the emerging field in their landmark survey Pre-train, Prompt, and Predict, demonstrating how diverse NLP tasks could be reformulated as cloze or prefix fill-in-the-blank templates. In January 2022, Jason Wei and colleagues at Google introduced Chain-of-Thought (CoT) Prompting, proving that prompting a large model to emit step-by-step intermediate reasoning before stating an answer unlocked dramatic gains on arithmetic and symbolic tasks, followed in May 2022 by Takeshi Kojima et al.’s discovery that simply appending “Let’s think step by step” (zero-shot CoT) triggered the same behavior. After ChatGPT’s release in November 2022, software-engineering researchers led by Jules White (February 2023) codified reusable “prompt patterns” modeled on object-oriented design patterns, while in 2023–2026 frameworks like DSPy (Khattab et al.) and automated prompt optimizers began replacing hand-tuned strings with programmatic compilation.

Central European Lineage

We found no documented link between this concept and the intellectual history or historical experience of Central Europe.

Why Now

Three milestones explain why prompt engineering became a focal point of both software practice and labor-market hype between 2020 and 2026. First, the launch of the GPT-3 API (June 2020) and ChatGPT (November 2022) turned natural language into the universal runtime interface for hundreds of millions of users, briefly sparking a 2023 gold rush of six-figure job postings for “prompt engineers” and an entire cottage industry of online courses selling secret templates. Second, in June 2024, Sander Schulhoff and dozens of collaborators published The Prompt Report, a massive taxonomy cataloging 58 distinct text-prompting techniques and 40 multimodal variants—bringing order to a sprawling empirical literature while acknowledging pervasive terminological confusion. Third, between 2024 and 2026, two forces reshaped the field simultaneously: on one side, instruction-tuned models and inference-time “reasoning” models (such as OpenAI’s o1/o3 series and DeepSeek-R1) internalized chain-of-thought search into their own post-training, rendering many manual prompting tricks redundant and shifting practitioner focus toward programmatic context engineering (deciding what data, tools, and memory state enter the context window); on the other side, the rise of autonomous LLM agents made prompt injection—where untrusted external text hijacks the model’s instructions—the number-one application security vulnerability in enterprise AI.

Explanatory Power

The concept makes visible a foundational truth of modern AI evaluation: a model’s observed behavior is never a property of the model weights alone, but a joint function of the model-plus-prompt pair. Without this lens, benchmark leaderboards and capability claims are nearly impossible to interpret. First, as Yao Lu and colleagues demonstrated in Fantastically Ordered Prompts (2021), simply changing the permutation order of four identical few-shot examples in a prompt can drive a model’s classification accuracy from state-of-the-art (nearly 90%) down to random guessing (50%)—proving that half of what looked like “model capability” was actually prompt-order luck. Second, Chain-of-Thought prompting (Wei et al., 2022) revealed a crucial distinction between latent competence and elicited performance: a model that fails flatly when forced to output a direct numerical answer in a single token pass often succeeds when prompted to write out intermediate scratchpad steps in its context window, because the generated tokens act as external working memory for subsequent attention layers. Third, the lens explains why two companies deploying the exact same frontier API endpoint experience wildly different reliability: the difference rarely lies in the model weights, but in how their system prompts, JSON schema constraints, and retrieved context chunks structure the probability distribution before the first output token is sampled.

Axes of Debate

Underappreciated Angle

The least examined aspect of prompt engineering is the fragile epistemic status of its own body of knowledge—which for its first several years resembled pre-scientific humoral medicine more than computer science. Thousands of viral posts and even peer-reviewed papers reported that telling a model “you are a world-class expert,” offering it a fictional monetary tip, or threatening it with punishment boosted benchmark scores—without testing whether those gains survived across different model families, different random seeds, or minor rephrasings of the evaluation dataset. When systematic replication studies finally tested these claims across diverse benchmarks, most persona prompts and emotional appeals showed erratic, near-zero average effect sizes, often helping on one subset of questions while hurting on another. Furthermore, as Lu et al. (2021) proved, a prompt configuration that is optimal for one checkpoint frequently degrades performance when transferred to a larger model or a rival lab’s API. The real intellectual breakthrough of prompt engineering is therefore the exact opposite of collecting “magic spells”: it is the realization that every prompt is an empirical hypothesis about a black-box distribution that must be versioned, ablated against a multi-hundred-example validation set, and automatically re-compiled whenever the underlying model changes. Those who run evaluation harnesses practice engineering; those who hoard prompt templates practice digital alchemy.