From GPT-2 to Today: How Prompt Engineering Evolved

Prompt engineering evolved from informal trial-and-error text manipulation around GPT-3 in 2020 to a structured discipline with named techniques, frameworks, and tools by 2026. The arc spans five phases: early few-shot experiments, the ChatGPT moment that brought the skill into mainstream awareness, the development of structured reasoning techniques, the rise of automated prompt optimisation, and the current shift toward context design.

The discipline did not emerge from a single paper or company. It grew from the overlap between research (few-shot learning, chain-of-thought reasoning, RAG), practitioner communities sharing prompt collections online, and the sudden public availability of powerful models that made good prompting immediately rewarding. By 2026, prompt engineering is no longer a niche trick — it is a baseline skill for anyone working with AI systems.

Before the term "prompt engineering" existed, researchers were already manipulating model inputs to elicit better outputs — they just did not call it that. Early transformer models like GPT-2 (2019, OpenAI) and BERT (2018, Google) were used through carefully chosen input text, but the practice was treated as part of data preprocessing, not a skill in its own right.

GPT-2, released in February 2019, was a 1.5-billion-parameter model that could complete text in surprisingly coherent ways. Researchers and early practitioners noticed that the phrasing of an input dramatically changed the quality of the completion — but there was no framework, no terminology, and no community built around this observation yet. Prompts were inputs, not engineering artifacts.

The modern history of prompt engineering effectively begins with GPT-3. In May 2020, OpenAI released GPT-3, a 175-billion-parameter model, alongside the landmark paper by Brown et al., "Language Models are Few-Shot Learners". The paper demonstrated that by including a few examples of the desired task directly in the prompt — without any weight updates to the model — performance on downstream tasks improved dramatically.

This was the seed of prompt engineering as a discipline. Researchers and developers realised that the same model could be turned into a translator, a summariser, a code generator, or a question-answering system simply by changing how the prompt was written. The model did not need retraining — it needed a better prompt. That insight reframed what a prompt was: not just an input, but a design artifact.

Brown et al. reported that few-shot performance scaled consistently with model size: the 175B GPT-3 model substantially outperformed smaller variants across every benchmark tested, establishing that scale and prompt-based learning were directly linked. This made the quality of the prompt a variable that practitioners — not just researchers — could control.

See Zero-Shot vs. Few-Shot: Which Approach Gets Better Results? for a practical guide to the technique GPT-3 made famous.

Between 2021 and early 2022, prompt crafting moved from research papers into practitioner communities. GitHub repositories with curated prompt collections appeared — "awesome-prompts" style lists that shared what worked for coding assistance, summarisation, and creative writing. Prompt collections, shared on Twitter and Reddit, became community assets. The Prompt Engineering Guide became one of the first dedicated references cataloguing techniques systematically.

The term "prompt engineering" began appearing more frequently in research papers, blog posts, and job descriptions through this period. OpenAI's InstructGPT paper (Ouyang et al., 2022) introduced RLHF-tuned models that responded far more reliably to natural-language instructions — making prompt quality even more consequential. By mid-2022, it was clear that this was a transferable skill, not just a researcher's curiosity.

The introduction of Chain-of-Thought (CoT) prompting in 2022 was the most significant technical development in the discipline's short history. Wei et al. (Google Brain) published "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", demonstrating that asking a model to reason step by step before answering dramatically improved performance on arithmetic, commonsense reasoning, and symbolic reasoning tasks. In one headline result, chain-of-thought prompting improved PaLM's accuracy on the GSM8K grade-school maths benchmark from 17.9% to 58% — a gain achieved purely by changing the prompt structure, with no additional model training. The implication was profound: the structure of the prompt could activate different reasoning behaviour — not just different facts.

Related techniques followed quickly. Zhou et al. introduced least-to-most prompting, which decomposed complex problems into a sequence of simpler sub-problems solved in order. These approaches turned prompt engineering from a formatting exercise into a tool for eliciting structured reasoning from models that had not been explicitly trained to reason that way. Prompting had become scaffolding for cognition.

For the full technique guide, see Chain-of-Thought Prompting: Make AI Show Its Reasoning and Prompt Chaining: How to Break Big Tasks Into Winning Steps.

The release of ChatGPT on November 30, 2022, changed the public profile of prompt engineering overnight. ChatGPT reached one million users within its first five days — confirmed by OpenAI CEO Sam Altman on Twitter in December 2022 — and 100 million monthly active users by January 2023, according to a UBS analysis cited by Reuters. Within days, millions of people were experimenting with prompts and discovering that their results varied enormously based on how they phrased requests. Tech media covered "prompt engineering" as a skill worth learning. The Oxford English Dictionary added "prompt" as a verb related to AI in 2023, and the word itself became a runner-up for word of the year in multiple rankings.

By early 2023, "prompt engineer" appeared as a job title with reported salaries of $175,000–$335,000 at companies including Anthropic, according to widely cited job postings. The role attracted significant media attention — Bloomberg, The Guardian, and The Atlantic all covered whether prompt engineering was a real career. The consensus at the time: it was a transitional role, part human-computer interface design, part subject-matter expertise, part quality assurance.

The popularisation of the phrase "prompt engineering" is sometimes attributed to various practitioners and commentators. Richard Socher, former Chief Scientist at Salesforce, is mentioned in some commentary as having helped frame the idea early. The Wikipedia article on prompt engineering provides a balanced overview of competing claims about the term's origins.

The release of GPT-4 in March 2023 expanded prompt engineering in two directions simultaneously: larger context windows (up to 128K tokens in later versions) and multimodal inputs. Practitioners could now include images in prompts alongside text, opening prompt engineering to visual tasks — describing images, comparing diagrams, annotating charts. Early Gemini models from Google and multimodal Claude versions from Anthropic followed within months.

The same year saw the formalisation of prompt engineering best practices. OpenAI published its official prompt engineering guide. Google Cloud released its own prompt engineering documentation. Independent authors codified frameworks — CRAFT, CO-STAR, SPECS, RISEN, TRACE — that gave practitioners repeatable templates for structuring prompts, reducing the reliance on trial and error.

These frameworks represented the maturation of prompt engineering from a personal skill into a teachable, shareable practice. See Which Prompt Framework Should You Use? for a guide to choosing between them, and Beyond Text: How to Prompt with Images for the multimodal dimension.

A striking development in 2023 was research showing that LLMs could optimise prompts as well as humans could. Zhou et al. published "Large Language Models Are Human-Level Prompt Engineers" (APE), demonstrating that an LLM tasked with generating and evaluating prompt candidates could match or exceed human-written prompts on benchmark tasks. Stanford's DSPy framework (2023) took this further — allowing developers to describe what a prompt should accomplish and letting the system optimise the wording automatically.

Simultaneously, Retrieval-Augmented Generation (RAG) — originally introduced by Lewis et al. at Meta in 2020 — became a central pattern in production AI systems. RAG injected retrieved documents directly into the prompt context, grounding model outputs in real, up-to-date sources rather than requiring prompts to contain all the necessary facts. This shifted the emphasis in prompt engineering from "how do I make the model know this?" to "how do I structure the context so the model uses this correctly?"

See RAG Explained: How to Ground AI Answers in Real Data and Self-Consistency Prompting: Let the AI Check Its Own Work for coverage of the key techniques from this period.

By 2024, a new framing began to displace the simple idea of "write a better prompt." Practitioners and researchers started referring to context engineering — the practice of orchestrating what goes into the full context window: the system prompt, retrieved documents, tool outputs, conversation history, and user input, all composed deliberately to guide model behaviour. The prompt was no longer a standalone artifact; it was one layer in a designed context.

Several developments accelerated this shift. Meta's Llama 3-class models (2024) made capable open-source LLMs available for private deployment, shifting some prompt engineering from cloud APIs to local infrastructure. Context windows grew to 1 million tokens or more (Gemini 1.5 Pro), making it practical to inject entire codebases, books, or document collections into a single prompt. Multi-agent frameworks like LangChain and AutoGen turned prompting into orchestration — one prompt triggers another model, which triggers a tool, which returns context to the next prompt.

As of 2026, research and commentary increasingly describe prompt engineering not as a niche job title, but as a fundamental literacy skill for knowledge workers who use AI tools. Academic papers frame structured prompting alongside reading, writing, and computation as a baseline competency for working with generative AI systems.

The role has split into two distinct tracks. The first is system and context design — the engineering of production AI systems where prompts form part of a larger architecture involving retrieval, agents, and evaluation pipelines. The second is everyday use — the ability to write clear, structured prompts that produce useful outputs without knowing the underlying architecture. Both tracks benefit from the same core principles: clear task specification, appropriate context, constraints, and output format.

What has not changed, despite more capable models and automated tools, is the fundamental principle: the clearer and more structured the input, the more reliable and useful the output. The techniques, terminology, and tooling have matured, but the core insight from the GPT-3 era remains true in 2026.

The table below summarises the key milestones from 2018 to 2026 — the events, papers, and model releases that shaped how prompt engineering evolved into its current form.

Each phase of prompt engineering's evolution left a lasting deposit in current practice. The GPT-3 era gave us the core insight that model behaviour is shaped by input structure — not just content. The Chain-of-Thought era gave us explicit reasoning scaffolds: step-by-step prompting, prompt chaining, and tree-of-thought approaches. The framework era gave us reusable templates that encode best practices without requiring each practitioner to discover them from scratch.

The RAG and context-design era gave us the understanding that prompts do not exist in isolation — they are composed with retrieved data, system instructions, and tool outputs to form a full context. And the automated-prompting era reminded us that the principles of good prompting are measurable: better-structured prompts produce better outputs in ways that can be evaluated and optimised systematically.