You opened a job board last week and the "prompt engineer" listings were gone. They got rebranded. In June 2025, two tweets did most of the work. Shopify CEO Tobi Lutke posted that he liked the term "context engineering" better than prompt engineering because it described the real skill: getting the right information into the model so the task is plausibly solvable. Andrej Karpathy reposted it the following week and added that prompt engineering had been collapsed in the popular mind into "typing things into a chatbot", while every industrial-strength LLM app was doing something far more delicate. Within a quarter, the rebrand was canonical inside teams that ship.
If you have been calling yourself a prompt engineer in 2026, hiring one, or paying for a course, this is the post that explains what changed and what to do about it. Prompts still matter. They are not where the leverage lives anymore. The leverage moved to the system around the prompt.
Quick takeaways: context engineering is the discipline of curating what enters the model's context window. The term was popularised by Tobi Lutke and Andrej Karpathy in June 2025 and formalised by Anthropic in September 2025. Context windows grew roughly 100x in two years, yet every model in Chroma's 18-model test degraded as input length grew. Bigger windows make curation more important, not less.
What "context engineering" actually means
Anthropic's Applied AI team defined the term on September 29, 2025 as "the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference." That is a long way of saying: you are the person who decides what the model sees on every turn, including which documents, which past messages, which tool outputs, and which instructions.
Compare that to the classic prompt-engineering frame: you are the person who writes the perfect sentence. Both still exist. The first one is now the job. The second one is a sub-task inside that job.
The receipts: why prompt engineering as a role is over
The market spoke before the discourse did. Indeed search interest for "prompt engineer" surged from 2 per million US searches in January 2023 to a peak of 144 per million in April 2023, then plateaued at 20 to 30 per million through 2025. Fortune reported in May 2025 that Indeed economist Allison Shrivastava called it "still definitely a good thing to have, but it's not an entire title." Microsoft's 2025 Work Trend Index, a 31,000-person survey across 31 countries, ranked "Prompt Engineer" second to last among roles companies planned to add in the next 12 to 18 months.
Demand for AI work did not fall. It migrated. AI and ML job postings on LinkedIn climbed sharply year over year in the first half of 2025. The work just stopped fitting under the "prompt engineer" header and started landing under titles like "Forward Deployed Engineer", "Applied AI Engineer", and at Contextual AI literally "Member of Technical Staff, Context Engineering". Sam Altman predicted in 2022 that prompt engineering as a job would not exist in five years. The clock has not quite run out. The trajectory is settled.
The physics: bigger windows made prompts less important
Here is the paradox almost nobody internalised. Context windows grew roughly 100x in two years. GPT-4 shipped in March 2023 at 8K tokens. As of May 2026, Claude Opus 4.6, 4.7, and 4.8 all expose a 1 million token window on the Anthropic API, Amazon Bedrock, and Vertex AI. Gemini 2.5 Pro sits at 1 million with 2 million flagged as coming. GPT-5 launched at around 400K, and OpenAI's GPT-5.5 release in April 2026 expanded that to 1 million via API.
Intuitively, that should make context engineering easier. Just dump everything in. The opposite happened. The Stanford and UC Berkeley "Lost in the Middle" paper (Liu et al., TACL 2024) showed a sharp U-shaped attention curve. Models hold the first few thousand tokens and the last few thousand. The middle goes effectively dark. Chroma's July 2025 "Context Rot" study tested 18 models including GPT-4.1, Claude 4, Gemini 2.5, and Qwen3 on trivially simple tasks. Every single one degraded as input length grew. Several models did better on a randomly shuffled haystack than on a logically coherent document.
Anthropic's own framing for this is the one to memorise: "Context must be treated as a finite resource with diminishing marginal returns." The bigger the window, the more your job becomes choosing what to leave out. That is context engineering in one sentence.
How the best AI products were actually built
If context engineering is the discipline, the proof is in what shipped. Three examples ride the same arc.
Cursor: the moat is the index
Cursor's coding agent is famous for feeling like it knows your codebase. The feeling is engineered. The pipeline has five steps: AST-aware chunking of every file (via tree-sitter), a Merkle tree of file hashes to detect what changed since last sync, an embedding pass, storage in Turbopuffer (a hybrid vector and full-text store), and a nearest-neighbour retrieval at query time. The prompt template at the end of that pipeline is a thin wrapper. The product is the index.
Claude Code: CLAUDE.md is the real prompt
Anthropic ships Claude Code with a convention called CLAUDE.md, a markdown file in the project root that the agent auto-loads into every session. It encodes architecture, code style, what to do, and especially what not to do. Practitioner consensus is that Claude reliably follows roughly 150 to 200 distinct instructions before the rule budget runs out. The system prompt eats a chunk of those. The rest is yours. The CLAUDE.md is, in effect, the prompt. The chat box is the cursor.
Perplexity: a multi-stage pipeline pretending to be a search box
Perplexity served roughly 780 million queries in May 2025, against its own growing knowledge index. Each query goes through parsing, embedding, hybrid retrieval (BM25 plus dense), a learned reranker, a structured prompt with pre-embedded citations, and a constrained generation step. The natural language "prompt" you type is the smallest layer of the stack.
The common thread is loud once you see it. None of these teams won by writing a smarter sentence. They won by building a smarter pipeline around the model.
The 7 skills a context engineer actually has
If you are trying to skill up for the work that replaced prompt engineering, here is the actual list. None of these are about clever phrasing.
- Retrieval design. Hybrid BM25 plus dense embeddings, reranking, query rewriting. The taste call is which step adds enough relevance to justify the latency cost.
- Chunking strategy. AST-aware for code, semantic for prose. Chunk size is the single most overlooked tuning knob in a RAG pipeline.
- Embedding model selection. When to use a vendor default like OpenAI text-embedding-3-large versus a specialist like Voyage or Cohere versus a fine-tuned in-house model.
- Memory architectures. Episodic for what happened, semantic for what is true, procedural for what to do. Compaction, summarisation, structured note-taking.
- Evaluation harnesses. Golden sets, regression evals, needle-in-haystack probes, end-to-end task evals. If you cannot measure context rot in your own product, you cannot fix it.
- Tool design. Anthropic's explicit recommendation: keep tool surfaces minimal and non-overlapping. The number of tools you expose is a context-budget decision.
- Agent loop engineering. When to spawn a sub-agent, when to compact, when to hand control back to the user. The hardest of the seven and the one most teams skip.
Here is what most takes on this topic miss
Almost every "prompt engineering is dead" post treats the shift as a vocabulary update. It is not. The economic shift is that the value of a single clever sentence is approaching zero, while the value of an indexed, versioned, evaluatable corpus is going up. A great prompt is a one-shot artefact. A great context system compounds every week you run it. That is why the only durable moats in AI products are data pipelines, retrieval quality, and memory. Not prompt libraries.
The honest counterargument
Simon Willison, who has written some of the clearest material on this stack, does not argue that prompt engineering died. He argues that "prompt engineering" as a public term got collapsed into "typing things into a chatbot", and that "context engineering" is a better label for the actual discipline. The underlying work never went away. It got renamed because the old name had become useless.
There is also a live disagreement worth flagging. Cognition's Walden Yan argued in June 2025 against multi-agent architectures: every action implies decisions, and parallel agents make conflicting decisions you cannot reconcile. Anthropic published its own multi-agent Research feature two days later, claiming a 90.2% lift over single-agent Claude Opus 4 on internal evals. Both positions are right inside their respective frames. Anyone telling you the multi-agent question is settled is selling you something.
What this means for your own work
If you build software, the practical translation is short. Stop optimising the prompt. Start optimising what gets into the prompt. Measure retrieval recall before you tweak phrasing. Compact aggressively. Keep CLAUDE.md or its equivalent under 200 lines. Treat context as the budget it is.
If you do not build software, the same idea has a quieter form. Every time you talk to an AI, you are doing context engineering by hand. The notes you brought, the screenshots you forwarded, the documents you uploaded: that is your context. The shape of your personal memory is the shape of the answers you can get. Models change every two months. Your context outlives them.
MemX is built on this idea. Your documents, voice notes, screenshots, and forwarded emails live in your account, indexed for retrieval, ready to be the context for whichever model you ask. The model changes. The context stays. memx.app
Key takeaway: prompt engineering is not so much dead as it is demoted. The job description that pays in 2026 is curating what the model sees, not what you type. The skill scales because every model upgrade leaves it intact.
01Who coined the term context engineering?
Tobi Lutke (Shopify CEO) popularised it in a tweet in mid-June 2025. Andrej Karpathy amplified it on June 25, 2025. The earliest serious technical articulation came from Walden Yan at Cognition on June 12, 2025. Anthropic formally adopted the term in an Applied AI engineering post on September 29, 2025.
02Is prompt engineering really dead?
As a job title in most companies, effectively yes. Microsoft's 2025 Work Trend Index ranked Prompt Engineer second to last among roles companies planned to add in the next 12 to 18 months. The underlying skill of writing clear instructions still matters. It just sits inside the broader context-engineering role now.
03Does a bigger context window solve the problem?
No. Chroma's July 2025 study showed every one of 18 tested models, including GPT-4.1 and Claude 4, degraded as input length grew. The 'Lost in the Middle' paper from Stanford and UC Berkeley showed a sharp U-shaped attention curve. Bigger windows make curation more important, not less.
04What is the single biggest mistake people make with LLM context?
Dumping everything in and hoping. Real teams curate. They retrieve only what is likely relevant, compact stale conversation, and keep instruction files short. The win is in what you leave out, not what you cram in.
05How do I start learning context engineering?
Read Anthropic's September 2025 engineering post, Cognition's 'Don't Build Multi-Agents', and Chroma's Context Rot study, in that order. Then build a small RAG pipeline end to end and instrument the retrieval step with a golden set. The discipline is mostly empirical.
06What is the difference between prompt engineering and context engineering?
Prompt engineering is writing the sentence you send to the model. Context engineering is choosing every token the model sees on a given turn: instructions, retrieved documents, prior messages, tool outputs, and memory. The prompt is one component inside the context.
