Prompt engineering is the practice of structuring what you send a language model so you get the output you want more often. For about two years it was treated as a job title. Since mid-2025 the fashionable position has been that it is finished, absorbed into the broader work of deciding what a model sees. That claim is half right, and the half that is wrong matters if you are deciding what to learn.
Updated September 2026. This area is moving quickly, so check the linked sources for the current position.
Where the argument came from
In June 2025 Andrej Karpathy publicly preferred the term context engineering, arguing that people associate prompts with short task descriptions while real applications are about filling the context window correctly. It was a remark about vocabulary. The framing spread quickly, and by 2026 a large amount of commentary had upgraded it from a preference about terminology into an obituary, usually without quoting anything that said the practice had stopped mattering.
The primary sources are more careful. Anthropic’s engineering post on the subject calls context engineering the natural progression of prompt work, not its replacement, and its 2026 practitioner guidance states plainly that prompting remains a fundamental building block within context engineering, alongside conversation history, files and system instructions. Our companion piece on context engineering covers the wider discipline in detail.
Sources disagree about the job market rather than the technique. Reported figures for postings and profiles carrying the title vary enormously depending on whether they count exact titles or any mention in a description, and the two produce opposite stories. What is not in dispute is the direction: the work is increasingly a component of engineering roles rather than a role of its own.
What the research actually covers
The field is better documented than the discourse suggests. The Prompt Report, a systematic survey first posted in June 2024 and revised through February 2025, assembled a vocabulary of 33 terms, a taxonomy of 58 text-only prompting techniques and a further 40 for other modalities, from a structured review of the literature. That is not a folk practice; it is a documented one.
What has changed is how much of it you need. Newer models follow instructions well enough that a lot of older scaffolding, elaborate role-play framing, repeated threats and incantations, no longer earns its place. Anthropic’s current advice describes prompting as converging with context engineering for its latest models: less scaffolding, more curation.
What prompt engineering still decides
Three things have not moved. First, ambiguity still produces variance: if you do not say what good looks like, the model will pick. Second, output format is still controlled by instruction and example, not by hope. Third, in an agent, the system prompt and the tool descriptions are prompts, and they run on every single step, so a sloppy sentence there is paid for thousands of times over.
The documentation is also clear that this is not where every problem should be solved. Anthropic’s overview tells you to define success criteria and build evaluations first, and notes that some failures are better fixed by changing model or approach than by rewriting the text.
What stopped working
A fair amount of early advice is now dead weight, and the published guidance quietly dropped it. Offering a model a tip, threatening it, or insisting it is a world-leading expert appears in none of the current vendor documentation. Long lists of forbidden behaviours tend to crowd out the instruction that actually matters. And copying a prompt that worked on one model into another rarely transfers cleanly, because each family’s own guidance differs in how it wants structure, reasoning and examples expressed.
The other casualty is the idea of a single magic wording. Because these systems are non-deterministic, a prompt that looks better on one run may be no better at all, which is why the published guidance keeps returning to evaluation rather than phrasing. If you cannot measure the change, you are collecting anecdotes.
6 habits worth keeping
These are the practices that survive the shift, drawn from the published guidance and the survey literature rather than from folklore.
- Define success before you write. Write down how you will tell a good output from a bad one, then test against it. Without that you are guessing.
- Be explicit, and say why. State the task, the audience, the constraints and the reason they matter. Models use the motivation, not just the instruction.
- Show one or two canonical examples, not twenty edge cases. Diverse, representative examples beat exhaustive ones.
- Ask for the reasoning when the task needs it, and skip it when it does not. Thinking steps cost tokens and latency and do not help every task.
- Give permission to refuse. Telling a model it may say it does not know reduces confident invention. Then still check the answer.
- Break long tasks into a chain. Several focused steps, each with a checkable output, beat one instruction carrying everything.
Prompting inside an agent is a different job
When a model runs in a loop with tools, the text you write is executed repeatedly and in situations you never saw. That changes the priorities. The system prompt wants to sit at the right altitude: specific enough to give real signal, general enough that it does not break the first time reality differs from your example. Tool descriptions become part of the prompt surface, and an ambiguous one produces the wrong tool call at step forty rather than a wrong sentence at step one. Error messages returned to the model are prompts too, and writing them for a reader who must decide what to do next is a real improvement you can ship in an afternoon.
What to learn instead of a job title
If you are choosing where to spend your time, spend it on evaluation, on retrieval and on the design of tools an AI agent will use. Writing a good instruction is a couple of hours to learn and a career to do well, but it is no longer the scarce part. Knowing whether your change helped is.
Common questions
Is prompt engineering dead in 2026? No. The standalone job title has largely disappeared into broader engineering roles, and heavy scaffolding matters less with newer models, but the underlying practice is now described as a building block inside context engineering.
What is the difference between prompt engineering and context engineering? Prompting is about the instructions you write. Context engineering is about everything the model sees at inference: instructions, tool definitions, retrieved documents, history and notes. The first is part of the second.
Do I still need to learn prompting techniques? Yes, but a small number of them. Clarity, examples, format control, staged tasks and permission to express uncertainty cover most practical gains.
Does prompt engineering still work on reasoning models? It works differently. Explicit step-by-step instructions matter less when a model reasons internally, while clear success criteria, constraints and output format matter as much as ever.
Is there a standard reference for prompting techniques? The Prompt Report, a systematic survey published on arXiv, catalogues 58 text-only techniques and 40 for other modalities with a shared vocabulary.
Sources and further reading
Where the figures and rules above come from, so you can check them:
- Best practices for prompt engineering, on convergence with context engineering: Anthropic
- Prompt engineering overview, on defining success criteria first: Claude platform docs
- Effective context engineering for AI agents: Anthropic
- The original preference for the term context engineering (June 2025): Andrej Karpathy
- The Prompt Report: A Systematic Survey of Prompting Techniques: Schulhoff et al., arXiv
Photo credits: Hacker typing on a Linux laptop full of stickers by User:Kowalski7cc, CC0, via Wikimedia Commons. Keyboard with fingers by Cangjie6, CC BY-SA 4.0, via Wikimedia Commons.
Prompting a model to write whole features has its own name and its own failure modes. See vibe coding.
Join the discussion