Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Prompt Engineering
Fundamentals

From GPT-2 to Today: How Prompt Engineering Evolved

·10 min readBy · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Prompt engineering evolved from informal trial-and-error around GPT-3 (2020) to a structured discipline with named frameworks, techniques, and tools by 2026, progressing through five phases: few-shot learning, chain-of-thought reasoning, the ChatGPT mainstream moment, automated optimization, and context design.

Key Takeaways

  • 2019–2020: GPT-2 and early transformers — prompts were inputs, not a discipline
  • 2020: GPT-3 and Brown et al. introduced few-shot prompting as a paradigm shift
  • 2022: Chain-of-Thought reasoning prompts turned prompting into a structured skill
  • Late 2022: ChatGPT brought prompt engineering into mainstream awareness and job postings
  • 2023: GPT-4, multimodal prompting, and frameworks formalised best practices
  • 2024–2026: Context design, automated prompting, and open-source LLMs redefined the field

Prompt engineering evolved from informal experiments around GPT-3 in 2020 to a structured discipline with named techniques and frameworks by 2026. This timeline traces the key breakthroughs, research papers, and turning points that made prompting a core skill.

From GPT-2 to Today: How Prompt Engineering Evolved

⚡ Quick Facts

  • ·GPT-3 (2020): 175-billion-parameter model introduced few-shot prompting as a paradigm shift
  • ·Chain-of-Thought (2022): Wei et al. showed that prompt structure could activate step-by-step reasoning; PaLM 540B went from 17.9% to 56.9% on GSM8K
  • ·ChatGPT (Nov 2022): Passed 1 million users in 5 days (per OpenAI CEO Sam Altman); an estimated 100 million monthly active users by January 2023 (UBS, via Reuters)
  • ·Job Market (2023): An Anthropic "Prompt Engineer and Librarian" posting advertised $175K–$335K (Fortune, March 2023)
  • ·GPT-4 (2023) & Frameworks: Image inputs and published frameworks (CO-STAR from GovTech Singapore, 2023; SPECS, RISEN) turned prompting into a teachable discipline
  • ·Context Design (2024–2026): Open-weights LLMs, 1M+ token context windows, and agent orchestration shifted focus from prompt tweaking to system-level context engineering

How Prompt Engineering Evolved: A Short Overview

📍 In One Sentence

Prompt engineering moved from trial-and-error phrasing tricks to a documented discipline of reusable patterns, evaluation and version control, as models became capable enough that structure mattered more than clever wording.

💬 In Plain Terms

It started as folklore — people swapping phrases that seemed to work. As the models got better, the magic words stopped mattering and the structure did: say who, what, and in what format. That shift is what turned a collection of tricks into a craft you can teach.

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 Prompt Engineering Had a Name (Pre-2020)

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 was announced in February 2019 as a 1.5-billion-parameter model; OpenAI released smaller versions first and published the full model in November 2019. It 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.

2020: GPT-3 and the Few-Shot Breakthrough

The modern history of prompt engineering effectively begins with GPT-3. In May 2020, OpenAI researchers published "Language Models are Few-Shot Learners" (Brown et al.), which described GPT-3, a 175-billion-parameter model. The paper showed that including a few examples of the desired task directly in the prompt — without any weight updates to the model — improved performance on many downstream tasks.

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 generally improved as model size grew, with the 175B-parameter GPT-3 the strongest variant overall. That linked scale and prompt-based learning, and it 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.

2021–Early 2022: From Prompt Tricks to a Recognised Skill

Between 2021 and early 2022, prompt crafting moved from research papers into practitioner communities. Researchers began describing the practice in print: Reynolds and McDonell's "Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm" (February 2021) argued that writing prompts for GPT-3 is a form of programming, and the survey "Pre-train, Prompt, and Predict" (Liu et al., July 2021) catalogued prompt-based methods across natural language processing. Practitioners meanwhile swapped working prompts in online communities.

OpenAI's InstructGPT paper (Ouyang et al., March 2022) introduced models fine-tuned with human feedback (RLHF) that followed natural-language instructions far more reliably than the base GPT-3, which made prompt quality even more consequential. By mid-2022 it was clear that prompting was a transferable skill, not just a researcher's curiosity.

2022: Chain-of-Thought and Reasoning Prompts

The introduction of Chain-of-Thought (CoT) prompting in 2022 was one of the most significant technical developments in the discipline's short history. Wei et al. (Google) published "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", showing that giving a model worked examples with intermediate reasoning steps improved performance on arithmetic, commonsense, and symbolic reasoning tasks. In one headline result, chain-of-thought prompting raised PaLM 540B's accuracy on the GSM8K grade-school maths benchmark from 17.9% to 56.9% — a gain achieved by changing the prompt, 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.

Late 2022–2023: The ChatGPT Moment and the Prompt Engineer Job Title

The release of ChatGPT on November 30, 2022, changed the public profile of prompt engineering overnight. OpenAI CEO Sam Altman said on Twitter in early December 2022 that ChatGPT had passed one million users in five days, and a UBS analysis reported by Reuters estimated about 100 million monthly active users by January 2023. Within days, millions of people were experimenting with prompts and discovering that results varied enormously with how a request was phrased, and technology media began covering prompt engineering as a skill worth learning.

By early 2023, "prompt engineer" had appeared as a job title: an Anthropic posting for a "Prompt Engineer and Librarian" advertised a salary range of $175,000–$335,000, as reported by Fortune in March 2023. Whether prompt engineering was a lasting career or a transitional role — part human-computer interface design, part subject-matter expertise, part quality assurance — was widely debated.

No single person is credited with coining the term. Research papers were already describing prompt-based methods by 2021 (see above), and the phrase spread widely after ChatGPT. The Wikipedia article on prompt engineering gives an overview of the history.

🔍 Did You Know

A UBS analysis reported by Reuters estimated that ChatGPT reached about 100 million monthly active users in January 2023, roughly two months after launch. Reuters, citing Sensor Tower data, put TikTok at about nine months and Instagram at about two and a half years to reach the same mark. This speed of adoption is why prompt engineering went from research concept to mainstream skill almost overnight.

2023: GPT-4, Multimodal Prompting and Frameworks

OpenAI announced GPT-4 in March 2023 with image input alongside text (initially a research preview), opening prompt engineering to visual tasks such as describing images, comparing diagrams, and annotating charts, and larger context windows (128K tokens with GPT-4 Turbo in November 2023) followed. Google's Gemini (December 2023) and Anthropic's Claude 3 models with vision (March 2024) added multimodal options from other vendors.

The same period saw the formalisation of prompt engineering best practices. OpenAI published its official prompt engineering guide. Google Cloud published its own prompt engineering guide. Independent authors and teams published frameworks — such as CRAFT, CO-STAR, SPECS, RISEN, and TRACE — that gave practitioners repeatable templates for structuring prompts, reducing the reliance on trial and error. The names, authors and publication dates of these frameworks vary by source, and none is an official standard.

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.

2023–2024: Automated Prompt Engineering and RAG

A striking development around 2023 was research showing that LLMs could optimise prompts as well as humans could. Zhou et al. posted "Large Language Models Are Human-Level Prompt Engineers" (APE) in November 2022; an LLM generated and scored candidate instructions and matched or beat human-written instructions on 19 of the 24 tasks they tested. The DSPy framework from Stanford researchers (Khattab et al., October 2023) took this further, letting developers describe what a prompt should accomplish and compiling the wording automatically.

Simultaneously, Retrieval-Augmented Generation (RAG) — introduced by Lewis et al. in 2020 while at Facebook AI Research (now Meta) — 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: Generate Multiple Answers, Pick the One That's Right for coverage of the key techniques from this period.

2024–2025: From Prompt Engineering to Context Design

During 2024–2025, a new framing began to displace the simple idea of "write a better prompt." 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 — moved into focus. The phrase itself gained traction in June 2025, when Shopify CEO Tobi Lütke and Andrej Karpathy publicly preferred "context engineering" over "prompt engineering", as documented by Simon Willison on 27 June 2025. 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 models (April 2024) made capable open-weights LLMs available for private deployment, shifting some prompt engineering from cloud APIs to local infrastructure. Google announced Gemini 1.5 Pro in February 2024 with a 1-million-token context window in limited preview, making it practical to place entire codebases, books, or document collections in 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.

2026 and Beyond: Prompt Engineering as a Core Literacy

By 2026, prompt engineering is increasingly described not as a niche job title but as a baseline literacy for knowledge workers who use AI tools: the ability to state a task, supply the right context, and specify the output you need.

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.

🔍 Pro Tip

The shift from "prompt engineering" to "context design" isn't just terminology — it changes what you optimize. Instead of tweaking the wording of your instruction, you design what goes into the context window: system prompt, retrieved documents, conversation history, tool outputs, and user input. The prompt is one layer, not the whole thing.

Timeline: Key Milestones in Prompt Engineering

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.

2018–2019

Milestone:
BERT (Google) and GPT-2 (OpenAI) released
Why It Matters:
Showed that transformer models could be steered by input phrasing, but no formal discipline existed yet

2020

Milestone:
GPT-3 and Brown et al., "Language Models are Few-Shot Learners"
Why It Matters:
Established few-shot prompting: rewriting the prompt changes the model's behaviour without retraining

2022 (Jan)

Milestone:
Chain-of-thought prompting (Wei et al.)
Why It Matters:
Showed that prompt structure could elicit step-by-step reasoning

2022 (Mar)

Milestone:
InstructGPT / RLHF (Ouyang et al., OpenAI)
Why It Matters:
Models trained to follow instructions made prompt quality far more consequential

2022 (Nov)

Milestone:
ChatGPT launch
Why It Matters:
Brought prompting to a mass audience; millions began experimenting within days

2023 (Q1)

Milestone:
"Prompt engineer" postings draw attention; an Anthropic posting advertises $175K–$335K
Why It Matters:
Made prompt engineering visible as a possible profession, though the role was widely debated

2023 (Mar)

Milestone:
GPT-4 release with image input
Why It Matters:
Extended prompting beyond text to images; 128K-token context followed with GPT-4 Turbo (Nov 2023)

2023

Milestone:
Frameworks such as CO-STAR (GovTech Singapore, 2023) and other published templates (CRAFT, SPECS, RISEN, TRACE); official guides from OpenAI and Google
Why It Matters:
Turned individual craft into reusable, teachable templates

2022–2023

Milestone:
APE (Zhou et al.) and DSPy (Khattab et al.): LLM-optimised prompts
Why It Matters:
Showed that LLMs can write and tune prompts, making automated prompt optimisation practical

2024

Milestone:
Llama 3 (Apr) and Gemini 1.5 Pro with a 1M-token context preview (Feb)
Why It Matters:
Open-weights models for private deployment and very long contexts moved focus to context engineering

2024–2025

Milestone:
Reasoning models: OpenAI o1 (Sep 2024), DeepSeek-R1 (Jan 2025), Claude 3.7 Sonnet extended thinking (Feb 2025)
Why It Matters:
Models reason step by step internally, so explicit chain-of-thought prompting matters less on them

2025 (Apr)

Milestone:
Llama 4 (mixture-of-experts); Llama 4 Scout supports a 10M-token context
Why It Matters:
Open-weights models adopt mixture-of-experts designs and very long contexts

2026

Milestone:
Context design and agent orchestration become standard practice
Why It Matters:
Prompting is one layer in a composed context; system-level thinking is expected

How the History Shapes Today's Best Practices

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.

FAQ: The Evolution of Prompt Engineering

When did prompt engineering start?

The practice began with GPT-3 in May 2020, when Brown et al. showed that examples placed in the prompt could steer a model without retraining. Researchers described it as "prompt programming" in early 2021, chain-of-thought prompting followed in January 2022, and ChatGPT (November 2022) made it a mass skill.

Who first coined the term "prompt engineering"?

The exact origin is debated, and no single person is credited with inventing the term. Research papers were describing prompt-based methods by 2021, and the phrase gained wider use through 2022 and after ChatGPT. The Wikipedia article on prompt engineering gives an overview of the history.

Why did prompt engineering explode in popularity after ChatGPT?

ChatGPT was among the first general-purpose AI models that millions of non-researchers could use immediately, for free, without writing code. The gap between a well-crafted prompt and a vague one was visible and immediately consequential — better prompts produced usably better outputs. That feedback loop, experienced simultaneously by millions of people, turned prompt engineering from a research concept into a mass skill.

How did research papers influence real-world prompting techniques?

The transfer was unusually fast for AI research. Chain-of-Thought prompting (Wei et al., 2022) went from academic paper to widely used practitioner technique within months, partly because it required no tooling — just a change in how you wrote the prompt. Few-shot prompting from the GPT-3 paper (Brown et al., 2020) was immediately adoptable by anyone with API access. The accessibility of the techniques accelerated their spread.

Is prompt engineering becoming less important as models improve?

No — more capable models respond better to well-structured prompts, not less. The gains from good prompting increase as the model becomes more capable of following precise instructions. What has changed is the level of prompt engineering required for simple tasks: conversational questions now require less crafting than they did in 2021. But for complex, production-grade outputs, structured prompting is usually one of the most effective levers, alongside retrieval, model choice and fine-tuning.

What is the difference between prompt engineering and context engineering?

Prompt engineering typically refers to designing the text input to a model to improve its output. Context engineering is a broader, more recent concept that refers to orchestrating everything in the model's context window: the system prompt, retrieved documents, conversation history, tool outputs, and user input — all composed deliberately. Context engineering treats the prompt as one component in a designed system, not a standalone artifact.

Will automated tools replace the need to understand prompt engineering?

Automated tools like DSPy can optimise prompt wording within defined objectives, but they require a human to specify what the objective is, what constraints apply, and how to evaluate success. Understanding prompt engineering principles remains necessary to use these tools effectively — and to diagnose when they produce the wrong outcome. Automation removes some of the manual iteration; it does not remove the need for structured thinking.

Is prompt engineering dead in 2026?

No. The discipline has shifted, not disappeared. As models grow more capable, the work moves from syntax tricks and formatting hacks to context design — structuring inputs, managing retrieval, and composing tool outputs. The job title "Prompt Engineer" is narrowing, but the underlying skill is embedded in every role that uses AI: developer, analyst, marketer, researcher. Clear task framing, relevant context, and a specified output format still determine how useful a model's answer is.

Do I need to learn prompt engineering if AI models keep improving?

Yes — but the focus shifts with each generation. Better models reduce the need for elaborate workarounds (special tokens, repetitive reinforcement, rigid formatting constraints) and increase the payoff for clear intent, structured context, and well-chosen examples. The fundamentals — role, context, format, constraints — remain stable across every model generation. Learning them now means the skill compounds rather than expires.

What is the difference between prompt engineering and fine-tuning?

Prompt engineering changes how you talk to a model without modifying its weights. Fine-tuning retrains a model on new data to change its behaviour permanently. Prompt engineering is faster, cheaper, and reversible — you can iterate in minutes. Fine-tuning is better when the target behaviour is consistent, high-volume, or impossible to describe reliably in a prompt. Most teams start with prompting and fine-tune only when prompting approaches a ceiling on their specific task.

Common Misconceptions About Prompt Engineering

❌ Prompt engineering is only about writing better sentences.

Why it hurts: This overlooks the structural and contextual dimensions. A prompt's effectiveness depends not just on word choice but on role assignment (assigning the model a persona), constraint specification, output format, and example selection — all structural elements that have nothing to do with grammar.

Fix: Think of prompt engineering as designing a system where the prompt is the interface. Invest in structure: assign roles ("You are a...", "Assume..."), specify constraints ("Do not...", "Must include..."), define output format, and provide examples. Structure often matters more than eloquence.

❌ Better models make prompt engineering irrelevant.

Why it hurts: A more capable model is like a more capable person — it responds *better* to clear instructions, not worse. The gains from good prompting can compound as model capability increases. What changes is the *kind* of prompting needed, not whether it's necessary.

Fix: Assume prompting will remain central to AI work. What evolves is the level of detail and scaffolding needed. With weaker models, you may need explicit step-by-step structure. With stronger models, a clear one-line instruction may suffice — but that directness is itself a prompt engineering choice.

❌ Automated prompt optimization tools will replace human prompt engineering.

Why it hurts: Automation tools like DSPy help optimize wording within defined objectives, but a human must still specify the objective, constraints, success criteria, and evaluation method. Automation removes iteration drudgery; it does not remove the need for structured thinking about what the model should do.

Fix: Use automation as a tool, not a replacement. Start with a well-structured prompt designed by a human who understands the task. Use tools like DSPy to refine and optimize it. The human judgement about task structure remains irreplaceable.

Sources

Apply these techniques with a local LLM or your own API keys — PromptQuorum works with any backend.

Try PromptQuorum free →

← Back to Prompt Engineering

MCP