Two years ago, Python developers were getting replaced by prompt engineers, at least in the tech Twitter space. The past year was all about vibe coding with tools like Cursor, Windsurf, Replit, and others. Now, the AI community has a new-found obsession: context engineering — the art and science of structuring everything an LLM needs to complete a task successfully.
If prompt engineering was about the clever, (mostly) one-liner instructions, context engineering is about writing the full screenplay. Andrej Karpathy, co-founder of OpenAI, also the person who called English the hottest programming language and made “tab tab tab” the default, is now all in favour of context engineering.
“+1 for ‘context engineering’ over ‘prompt engineering’,” Karpathy said in a post on X. He added that he does not want to coin a new term for it now that it has already caught on with developers.
“In every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step,” said Karpathy.
From Prompts to Context
Prompt engineering gave us the early magic of ChatGPT — coaxing the model into doing our bidding with clever phrasings. But as applications get complex, that approach hits a wall.
“People associate prompts with short task descriptions,” Karpathy explained. “But apps build contexts — meticulously — for LLMs to solve their custom tasks.”
Unlike prompt engineering, which focuses on how to phrase a task, context engineering is about ensuring the task is possible to solve in the first place. That might mean retrieving relevant documents using RAG, summarising a long conversation to preserve state, injecting structured knowledge, or supplying tools that let the model take action in the world.
Many developers are realising that when LLMs fail, it’s not because the model is broken — it’s because the system around it didn’t set it up for success. The context was insufficient, disorganised, or simply wrong. And like humans, LLMs respond differently depending on how you talk to them.
A poorly structured JSON blob might confuse a model where a crisp natural language instruction would succeed.
The shift to context engineering is not just semantic. It’s structural. Where prompt engineering ends at crafting a sentence, context engineering begins with designing full systems, ones that bring in memory, history, retrieval, tools, and clean data — all optimised for an AI model that isn’t psychic.
“Context engineering is 10x better than prompt engineering and 100x better than vibe coding.” That’s how Austen Allred, founder of BloomTech, summed up the shift in how developers are thinking about building LLMs.
Context is King
Sebastian Raschka, founder of RAIR Lab, captured the division well: prompt engineering is often user-facing, while context engineering is developer-facing. It requires building pipelines that bring in context from user history, prior interactions, tool calls, and internal databases — all in a format that’s easily digestible by a Transformer-based system.
Context engineering doesn’t just mean “adding more stuff” to your prompt. It means curating, compressing, and sequencing the right inputs at the right time. It’s a system, not a sentence.
Harrison Chase, CEO and co-founder of LangChain, said that this is precisely why frameworks like LangGraph are gaining traction. Designed to give developers fine-grained control over what goes into the model, what steps run beforehand, and where outputs are stored, LangGraph embraces the philosophy that context engineering is central to any serious agent framework.
While older abstractions often hide this complexity in the name of ease of use, LangGraph puts context back into the developer’s hands, where it belongs.
Context engineering has organisational implications beyond the technical realm. Ethan Mollick, associate professor at The Wharton School, noted that it’s not just about crafting a useful LLM input — it’s about encoding how your company works.
That includes the structure of reports, the tone of communication, and the internal processes that define your business logic. In that sense, context engineering is as much about culture as it is about code.
Karpathy made a broader point that often gets lost in the discussion: context engineering is just one piece of a growing software stack built around LLMs. It coexists with problem decomposition, memory management, UI/UX flows, verification steps, and orchestrating multiple LLM calls.
Calling all of that a “ChatGPT wrapper,” he said, is “really, really wrong.” It’s not a wrapper — it’s a new paradigm of software altogether.
And perhaps that’s why the term “vibe coding” is now being used tongue-in-cheek. In the early days of LLM experimentation, developers often relied on intuition and repetition, tweaking wording endlessly until they got something that felt right.
But intuition doesn’t scale; structure does. What works in a playground doesn’t hold up in production. Tobi Lütke, CEO of Shopify, said it best: “It describes the core skill better — the art of providing all the context for the task to be plausibly solvable by the LLM.”
I really like the term “context engineering” over prompt engineering.
It describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM.
— tobi lutke (@tobi) June 19, 2025
That word — plausibly — carries weight. AI models don’t have intent or judgment. They’re not reasoning from first principles. They’re guessing the next word based on everything you’ve told them so far. And if you haven’t told them the right things, or told them in the wrong format, your clever prompt won’t save you.
Context engineering is not just the new vibe; it’s the new software architecture.
