What It Is
Terminology · LLM Engineering · Language Models · Benchmarking · AI · ExplainerTratopedia · 15 Aug 2026
First of five · the innermost layer
The Job Title Died. The Sensitivity Did Not.
Prompt engineering is writing the instruction so the model does what you meant. It got a dictionary entry, a job title, and then an obituary: the trade press called the role obsolete within two years of inventing it. But the measurements that made it a discipline in the first place have not been reported as fixed. Reordering the examples in a prompt has moved accuracy by more than forty points. Changing only the formatting has moved it by more than seventy. That sensitivity persists with bigger models, more examples and instruction tuning. What ended was a profession, not a phenomenon — and the practice did not stop so much as move, from the whole of how you use a model to one component inside something larger.
Three definitions, and the one word that moved All three agree on the practice. Only one of them calls it a profession.
- 40+accuracy points that reordering the examples in a prompt has been found to move
- 76accuracy points across formatting changes alone, in few-shot settings
- 8chain-of-thought exemplars that took a 540B model to state of the art on GSM8K
- 2years from the job title being the hottest in AI to the trade press calling it obsolete
Oxford English Dictionary · 2025
“…and the discipline or profession concerned with this”
- In full: the action or process of formulating and refining prompts for an artificial intelligence program, algorithm, etc., in order to optimize its output or to achieve a desired outcome; the discipline or profession concerned with this.
- That final clause is the half that moved. In 2023 “prompt” was runner-up for Oxford's word of the year.
Anthropic · Sep 2025
“Methods for writing and organizing LLM instructions”
- No profession, no discipline. Just the writing.
- The same post calls context engineering prompt engineering's natural progression — not its replacement.
Wikipedia · at capture
“Structuring natural language inputs to produce specified outputs”
- It puts context engineering beside it as “the related area… focused on non-prompt and prompt contexts” — system instructions, metadata, API tools, tokens.
- Two sources from opposite directions, one shape: the prompt stayed, everything around it got a name of its own.
How firmly each part stands Four papers were read at source. The numbers everyone quotes were not.
| How firm | What | On what |
|---|---|---|
| Read at source | Eight chain-of-thought exemplars took a 540B-parameter model to state-of-the-art accuracy on GSM8K, surpassing even finetuned GPT-3 with a verifier | Wei et al., arXiv 2201.11903, submitted 28 Jan 2022. Abstract read at source |
| Read at source | GPT-3: 175 billion parameters, ten times any previous non-sparse model; few-shot use “sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches” | Brown et al., arXiv 2005.14165, submitted 28 May 2020. Abstract read at source |
| Read at source | Ten NLP tasks recast as question answering over a context, one model, no task-specific parameters | McCann et al., arXiv 1806.08730, submitted 20 Jun 2018. Abstract read at source |
| Read at source | The OED's definition, including “the discipline or profession concerned with this” | Quoted by Wikipedia from Oxford University Press, 2025. The dictionary entry itself was not opened |
| Via Wikipedia | More than 40 accuracy points from reordering examples; up to 76 from formatting changes alone in few-shot settings | Wikipedia, citing studies. Those studies were not read here — the magnitude is what is used, not any single result |
| Via Wikipedia | The sensitivity persists despite larger models, more few-shot examples, and instruction tuning | Wikipedia. This is the claim the whole argument turns on, and it is the one carried by the weakest source here |
| Via Wikipedia | In-context learning is an emergent property of scale, and temporary — unlike training or fine-tuning, it leaves nothing behind | Wikipedia |
| Headline only | The role was declared obsolete: AI Prompt Engineering is Dead (Mar 2024); The Hottest AI Job of 2023 Is Already Obsolete (Apr 2025) | IEEE Spectrum and the Wall Street Journal. Only the headlines were seen; neither article was read, and they are used here only as evidence that the genre exists |
| Our reading | The scope of the word shrinks at every step while the phenomenon does not | Assembled across all of the above. No source consulted presents this progression |
Timeline
What the word meant, each time it moved Not a list of events — a list of meanings. Every date is a publication, not a change in how people spoke.
- Jun 2018A prompt is a question that picks a task. decaNLP recasts ten NLP tasks as question answering over a context — “What is the sentiment?”, “Translate this to German” — with one model and no task-specific parameters.
- 2019The prompt replaces fine-tuning. GPT-2 performs downstream tasks zero-shot, “without any parameter or architecture modification”. The instruction is now the whole specification.
- May 2020From asking to teaching. GPT-3, at 175 billion parameters, makes few-shot prompting work: the prompt now carries worked examples, and prompting becomes in-context learning.
- Jan 2022From the answer to the method. Chain of thought: eight exemplars showing reasoning, not results, take a 540B model to state of the art on GSM8K. Note the date — this is ten months before ChatGPT.
- 2022And back to a phrase. Appending “Let's think step-by-step” is found to work with no demonstrations at all. The scope collapses from a technique to a sentence.
- 2023Peak: it becomes a profession. An OED headword whose definition ends “the discipline or profession concerned with this”. “Prompt” is runner-up for Oxford's word of the year. Companies hire prompt engineers.
- 2024–25The profession deflates. IEEE Spectrum runs “AI Prompt Engineering is Dead” in March 2024; the Wall Street Journal calls the 2023 job obsolete in April 2025. Wikipedia notes the title became less common — models write better prompts than people, and everyone else was trained in it.
- Sep 2025Relocated, not removed. Anthropic calls context engineering the natural progression of prompt engineering. The prompt stays; everything around it — tools, memory, message history — gets a name of its own, and the prompt becomes one component of it.
The Argument
Is it dead? Two bodies of evidence, pointing opposite ways, about two different things.
The case for the obituary is real. The job title appeared and vanished inside two years, and Wikipedia gives two reasons rather than one: models now write better prompts than people do, and ordinary employees were trained in the skill, so a specialist role had nothing left to be specialist about. Both are plausible and neither is quantified anywhere consulted. The case against is that the measurements never moved. If output quality swings by forty accuracy points when you reorder the same examples, and by seventy-six when you change nothing but the formatting, then something is still there to engineer — and the finding is explicitly that this persists with larger models, more examples and instruction tuning. That is the load-bearing claim of this whole page, and it is carried by its weakest source: Wikipedia's citation of studies this session did not read. It is marked that way in the table above and it should be held loosely. The two bodies of evidence are not actually in conflict, because they are about different things. One is about employment. The other is about models. A practice can stop being a career and go on being necessary — which is roughly what both Anthropic and Wikipedia describe, from opposite directions, when they put the prompt inside something larger rather than throwing it away.
- 2 yrfrom hottest job in AI to declared obsolete, by the trade press's own headlines
- 40 ptsmoved by reordering the same examples. Nothing consulted says this was fixed
- 2different subjects: one body of evidence is about employment, the other about models
What Others Add
What the techniques actually were The named methods, and what each one added to the one before it.
| Technique | What it does | What it added |
|---|---|---|
| Few-shot | Put worked examples in the prompt | The model can learn a task from the prompt alone. That ability is emergent with scale, and temporary — unlike fine-tuning, nothing is left behind afterwards |
| Chain of thought | Examples that show the reasoning steps, not just the answer | Shapes the model's process. Eight exemplars, a 540B model, state of the art on GSM8K |
| Zero-shot CoT | Append “Let's think step-by-step” and drop the examples | Showed the effect did not need the demonstrations at all — one phrase reached it |
| Self-consistency | Run several chains of thought, take the most common conclusion | Turns one sampled answer into a vote |
| Tree of thought | Several lines of reasoning in parallel, with backtracking | Search, rather than a single pass — breadth-first, depth-first or beam |
And where it went Three places the same practice turns up under other names.
- 1Into the system promptAnthropic's guidance is to write at the right altitude — between hardcoded brittle if-else logic and vague guidance that assumes shared context. Aim for the minimal set that fully specifies the behaviour, noting that minimal does not mean short.
- 2Into tool definitions“Tool definitions and specifications should be given just as much prompt engineering attention as your overall prompts.” On SWE-bench, Anthropic reports spending more time optimising the tools than the prompt.
- 3Into a layer of something biggerContext engineering takes over the surrounding job — what the model can see and remember at each step. Prompt engineering does not disappear from it; it becomes the part that writes the instruction, once per step instead of once per task.
One detail from the papers is worth keeping, because it is usually lost in the retelling. Chain of thought was submitted on 28 January 2022 — ten months before ChatGPT was released. The technique most often held up as prompt engineering's high point arrived before the audience did, which is a useful corrective to a story that otherwise starts with a chat box. A second detail: Wikipedia describes the chain-of-thought result as performing “comparably with task-specific fine-tuned models”. The paper's own abstract says state of the art, and surpassing finetuned GPT-3 with a verifier. Both sentences are about the same experiment; one is simply weaker than what was claimed. This page uses the paper's wording, and the difference is the ordinary reason for following a summary back to its source.
- 10 mochain of thought predates ChatGPT's release by ten months
- SOTAwhat the paper claims, against “comparably” in the summary of it
- 175Bparameters in GPT-3 — ten times any previous non-sparse model
Conclusion
What to take away Three things, and the first is about how to read the obituaries.
A job title ended. That is not the same as a practice ending
Two headlines declared the role obsolete. Neither is about whether wording still moves output — and the measurements that say it does are not reported as fixed anywhere consulted. Read “prompt engineering is dead” as a claim about hiring, because that is what the evidence behind it covers.
The word's territory shrank at every single step
A task selector in 2018, a replacement for fine-tuning in 2019, a way of teaching by example in 2020, a way of shaping reasoning in 2022, then a phrase, then a profession, then one component inside context engineering. This progression is assembled here and appears in no single source — but every step in it was read at its own publication.
The strongest claim here rests on the weakest source
That the sensitivity persists with scale, more examples and instruction tuning is what makes the whole argument work — and it comes from Wikipedia citing studies this session did not read. It is marked as such in the table. If you are going to check one thing on this page, check that.