What Is Context Engineering? The Skill That Replaced Prompt Engineering

For two years, «prompt engineering» was the skill everyone rushed to learn: the art of wording the perfect instruction for ChatGPT. In 2026 the phrase you hear instead is context engineering — and Gartner has gone as far as saying «context engineering is in, prompt engineering is out». If you keep seeing the term and want to know what it actually means, without the hype, this article is for you.

We build AI chatbots for a living, so we care about this for a very practical reason: it's the difference between an AI that makes things up and one that answers with your real information. We'll explain what context engineering is, where the term came from, how it differs from prompt engineering (and whether prompt engineering is really «dead»), the pieces that make up «context», and what all of it means for a business chatbot. With the real sources, not the viral threads.

What context engineering is

Context engineering is the practice of assembling the right information around an AI model so it answers correctly. Not the wording of a single question, but everything the model can «see» when it responds: your documents, your data, the instructions, previous messages, the tools it can use. Anthropic, which published the reference guide on the topic, defines it as «the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference».

The key idea is that the language model itself — GPT, Claude, Gemini — is the same one everyone has access to. What makes it answer your customer's question correctly isn't a smarter model; it's the context you place in front of it. Give it the right paragraph from your returns policy and it answers precisely. Give it nothing and it guesses. Context engineering is the discipline of getting that right, on purpose and at scale.

Where the term came from

The phrase went mainstream in mid-2025. Tobi Lütke, the CEO of Shopify, popularized it in a June 2025 post, describing context engineering as the art of providing all the context needed for a task to be plausibly solvable by the model. A week later, Andrej Karpathy (a founding member of OpenAI) endorsed it publicly, and his framing is the one most people quote:

“+1 for «context engineering» over «prompt engineering». People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When 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.”

— Andrej Karpathy

By September 2025, Anthropic had formalized it in an engineering guide, and the analysts followed: Gartner declared that «context engineering is in, and prompt engineering is out», calling it a breakout AI capability and predicting it will show up in the majority of AI tools within a few years. In other words: this isn't a random buzzword a marketing team invented. It came from the people actually building production AI, and the industry agreed it named something real.

Context engineering vs prompt engineering

Here's the honest version, because the viral headlines oversell it. Prompt engineering is about how you word the instruction — phrasing the question or the system prompt well. Context engineering is about what information the model has in front of it when it answers — the instruction being just one piece of it.

Is prompt engineering dead, then? No — and anyone telling you it is flat-out is selling something. A well-written system prompt is still essential; it's just no longer the whole job. Prompt engineering is now one layer inside context engineering — the instructions layer. What changed is the realization that a perfectly worded prompt with the wrong information in front of it still gives a wrong answer. The wording matters; the information matters more.

The pieces of «context»

When Anthropic breaks «context» down, it's not one thing — it's a handful of ingredients that all compete for the same limited space. These are the pieces:

• System instructions — who the assistant is, what it does and the rules it follows (this is where prompt engineering lives). We cover this one in depth in the system prompt guide.
• Retrieved data — the specific pieces of your information pulled in for this question: a paragraph from a document, a row from a catalog. This is what retrieval (RAG) does — with newer variants like PixelRAG changing how it works.
• Tools — the actions the model can take (look up an order, check stock), each one described so it knows when to use it. This is the piece that turns a chatbot into an AI agent.
• Examples — a few canonical samples of the behavior you want, worth more than a long list of rules.
• Message history — the earlier turns of the conversation, kept (or trimmed) so the model remembers what was already said.
• Memory — notes persisted outside the conversation, so the assistant doesn't start from zero every time.

The craft is that the context window is finite. You can't dump everything in — too much context is as harmful as too little (a problem researchers call context rot). Anthropic sums up the goal as finding «the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome». Not the most information; the right information.

Why context engineering stops AI from hallucinating

Here's the part that matters for anyone running an AI on real data. A model «hallucinates» — makes up a confident, wrong answer — mostly when it doesn't have the real answer in front of it and fills the gap with a plausible guess. Give it the actual returns policy in its context and it doesn't need to invent one.

That's why context engineering and reliability are the same conversation. Grounding the model in retrieved, real information is the single most effective way to cut hallucinations. We wrote a whole piece on the failure modes in why your chatbot makes things up — almost every fix in it is, underneath, a context-engineering fix: the bot answered badly because the right information wasn't in its context when it needed it.

What it means for a business chatbot

Put all of this together and you get a useful reframing: a business chatbot is a context-engineering problem, not a model problem. Every serious chatbot runs on the same handful of frontier models. The one that answers your customers well isn't the one with a secret smarter AI — it's the one that puts the right slice of your business in front of the model at the right moment.

And that's exactly where most chatbots quietly fail. They dump a PDF into the model and hope; they retrieve text by rough similarity and miss the exact price or size; they forget the earlier messages. Those are all context-engineering mistakes. The difference between a chatbot that frustrates and one that helps is almost never the model — it's whether the context around it was engineered well.

How Bravos does the context engineering for you

The good news for a business owner is that you don't have to become a context engineer. That's the platform's job. With Bravos AI you provide the knowledge — your documents, your website, your catalog — and the platform assembles the context for every question so the bot answers grounded in your real data, not guesses. In practice that means three things you get without configuring anything:

• The right information retrieved per question, so the model sees the relevant part of your content instead of everything or nothing.
• Exact answers on structured data: for catalogs and anything with prices, sizes, stock or references, it filters by the real value instead of «something similar». We explain why that matters in chatbot for product catalogs.
• Your instructions and the conversation kept in context, so the tone and the rules you set are respected turn after turn.

You don't write prompts or wire pipelines. You upload your data; the context engineering happens under the hood. That's the whole point of a good platform: it turns «the skill everyone's talking about» into something you get by default. And if you already work with Claude or ChatGPT, you can inspect and adjust that context yourself in plain language — the bot's knowledge and its instructions — through the Bravos MCP connector (new to the term? here's what an MCP is).

In short

Context engineering is the practice of assembling the right information — instructions, your data, tools, memory — around an AI model so it answers correctly instead of guessing. It replaced «prompt engineering» as the headline skill not because wording stopped mattering, but because the industry realized the information around the model matters more than the phrasing of a single prompt.

For a business, the takeaway is simple: a chatbot is only as good as the context it's given. Get that right and it answers with your truth; get it wrong and it invents. And the practical way to get it right without becoming an engineer is to use a platform that does the context engineering for you.

Frequently asked questions

Sources

• Anthropic — Effective context engineering for AI agents: anthropic.com — the reference technical guide (Sept 2025), source of the definition and the components of context.
• Andrej Karpathy on context engineering: x.com/karpathy — the post that popularized the term (June 2025).
• Gartner — Context engineering vs prompt engineering: gartner.com — the analyst view («context engineering is in, prompt engineering is out»).