Prompt Engineering

Prompt engineering is the process of structuring natural-language inputs — prompts — to produce specified outputs from a generative AI model.¹ The Oxford English Dictionary gives it as “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.” In 2023, prompt was runner-up for Oxford’s word of the year.

This page exists mostly to answer a question that gets asked in a misleading form. The short version: the job title is largely gone, the term has been superseded as the name of the discipline, and the practice is alive and sitting inside context engineering. Those three fates are different, and collapsing them is where “prompt engineering is dead” goes wrong. The long answer is below.

The techniques

A 2024 survey found over 50 distinct text-based prompting techniques, 40 multimodal variants, and a vocabulary of 33 terms — which the article reads, correctly, as evidence that the field lacks standardized terminology.¹ The durable ones:

• Multi-shot / few-shot — include examples for the model to learn from in-context, e.g. completing maison → house, chat → cat, chien →. This is in-context learning, and it survives intact as one of the components of context engineering.
• Chain-of-thought (CoT) — proposed by Google researchers in 2022; induces the model to answer a multi-step problem with intermediate reasoning steps. Applied to PaLM (540B) it reached state-of-the-art on mathematical reasoning benchmarks. Originally few-shot, it was later found to work zero-shot by simply appending “Let’s think step-by-step.”
• Self-consistency — run several CoT rollouts and take the most common conclusion.
• Tree-of-thought — generalizes CoT to multiple parallel reasoning lines with backtracking, using breadth-first, depth-first, or beam search.
• Role assignment — instruct the model to adopt a persona.
• Retrieval-augmented generation (RAG) — retrieve from a document collection at query time so the model answers from specified sources, enabling domain-specific and current information without retraining. GraphRAG (Microsoft Research) extends this with knowledge graphs so the model can connect disparate facts and summarize semantic concepts across large datasets.

CoT’s benefits are narrower than its reputation. A meta-analysis of 100+ studies found substantial gains concentrated in mathematical, logical, and symbolic reasoning, with much smaller improvements elsewhere — and on cognitive- psychology tasks where deliberation impairs human performance, CoT reduced model accuracy.¹ It is a tool for a class of problems, not a general-purpose accuracy switch.

This wiki covers several of these from the practitioner’s angle already: zero/one/few-shot and role prompting in GAIL Module 4, and CoT and ReAct as reasoning-loop techniques in GAIL Module 5. The ReAct pattern is where a prompting technique becomes an architecture.

Why prompts are brittle

The reason prompt engineering never stabilized into reliable practice is that model behavior is extremely sensitive to things that carry no meaning.¹

• Reordering the examples in a prompt has produced accuracy shifts of more than 40 percentage points.
• Formatting changes alone have produced swings of up to 76 accuracy points in few-shot settings.
• Linguistic features — morphology, syntax, lexico-semantic choices — measurably change task performance. Clausal syntax, for instance, improves consistency and reduces uncertainty in knowledge retrieval.
• The sensitivity persists despite larger models, more few-shot examples, or instruction tuning.

Two structural limitations follow, and they’re why the discipline stayed a craft: effective strategies are model-specific — a technique that helps one model can hurt another — and prompts are brittle, in that minor changes to phrasing, punctuation, or word order produce dramatically different outputs at identical semantic intent. Together these make durable, transferable prompt patterns hard to establish.¹

Proposed responses are evaluative rather than prescriptive: FormatSpread evaluates a range of plausible prompt formats to report a performance interval instead of a point estimate, and PromptEval estimates performance distributions across diverse prompts to give robust metrics like quantiles. Both amount to admitting that a single measured prompt score is not a real number.

In-context learning

A model’s ability to learn temporarily from its prompt is in-context learning — an emergent property of scale, whose efficacy increases at a different rate in larger models than smaller ones. Unlike training and fine-tuning, which change parameters durably, in-context learning is temporary; training a model to do it can be seen as meta-learning, “learning to learn.”¹ This is the mechanism every prompting technique on this page ultimately exploits.

Automating it

The response to brittleness has been to stop doing it by hand.¹

• Automatic prompt engineer (APE) — one LLM beam-searches over prompts for another: show it input/output pairs, ask for instructions that would produce those outputs, score each by the log-probability of the outputs, then refine the best.
• Auto-CoT — vectorize a question library, cluster it, pick diverse centroid-proximate questions, run zero-shot CoT on each, and use the results as few-shot demonstrations.
• MIPRO — optimizes instructions and few-shot demonstrations across multi-stage programs.
• GEPA — combines LLM-based execution-trace analysis with Pareto evolutionary search, reporting 10% gains over reinforcement learning with 35× fewer rollouts.
• Soft prompting (prefix-tuning, prompt tuning) — gradient descent over floating-point vectors rather than tokens, maximizing output log-likelihood.
• DSPy and similar frameworks expose these optimizers as programmatic pipeline components.

This branch is growing, not shrinking, and it complicates any simple “deprecated” story: the parts of prompt engineering that could be handed to a compiler are being handed to one.

Prompt injection

Prompt injection is a security exploit in which an adversary crafts input that looks legitimate but causes unintended model behavior. It works because models cannot reliably distinguish developer-defined instructions from user input, which lets an attacker bypass safeguards and steer behavior.¹

This is the part of the field that has become more important rather than less. Every technique on this page that widens what enters the context window — RAG, tool output, retrieved documents, agent-to-agent messages — widens the injection surface with it. It’s the security cost of context engineering, and it belongs on the same ledger.

Is it deprecated?

Not as a practice. Three different things travel under this name, and they have had three different fates.

The job title is effectively obsolete. Employees titled “prompt engineer” were hired across industries during the 2020s AI boom, but the title has become uncommon — models now produce better prompts than humans, and prompting is taught as general corporate training rather than staffed as a specialty. The Wall Street Journal put it bluntly in 2025: the job “was one of the hottest in 2023, but has become obsolete due to models that better intuit user intent and to company trainings.”¹

The term has been superseded as the name of the discipline. This is what Lütke and Karpathy were arguing in June 2025 — not that prompting stopped mattering, but that the word misdescribes the work. Karpathy’s objection is specifically lexical: “people associate prompts with short task descriptions you’d give an LLM in your day-to-day use,” while the actual job in a production app is filling the whole context window.² The complaint is that the term is too small, not that the activity is worthless.

The practice is alive, and now a component of something larger. Anthropic frames context engineering as “the natural progression of prompt engineering,” and its own guidance devotes sections to system prompts — clarity, the right altitude, sectioning with XML tags or Markdown headers — and to few-shot examples, which it “continue[s] to strongly advise.”³ You cannot do context engineering without doing prompt engineering; the prompt is simply no longer the whole of what you engineer.

What genuinely eroded is narrower than the headline suggests: manual, trial-and-error prompt tinkering for everyday tasks. As models improved at interpreting intent directly, the marginal value of elaborate prompt construction fell, which is what raises the question about the field’s long-term viability as a standalone discipline.¹ Note the qualifier — standalone. The verdict in the sources is absorption, not deletion.

What did not erode: system prompt design (harder now, because agents run longer), few-shot example curation, automated prompt optimization (actively advancing), and prompt injection defense (more urgent every year). And Wikipedia’s own lead now introduces context engineering in its first paragraph as the related discipline covering non-prompt context — the encyclopedia has already absorbed the reframing.¹

The honest summary is that prompt engineering is in the position of “webmaster” circa 2005: the title is gone, the word sounds dated, and every task it named is still being done by someone with a different title.

Footnotes

1. “Prompt engineering”, Wikipedia, revision 1373316427 (2026-09-05). Local excerpt: references/wikipedia-prompt-engineering.md. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

2. Tobi Lütke (2025-06-19) and Andrej Karpathy (2025-06-25), posts on X. Local copy: references/context-engineering-coinage-tweets.md. ↩

3. Anthropic Applied AI team, “Effective context engineering for AI agents”, 2025-09-29. Local copy: references/anthropic-effective-context-engineering.md. ↩