What Is Context Engineering?
Why the bottleneck shifted from prompts to what the model sees
Prompt engineering had a good run. The difference between a useful AI response and a useless one used to come down to how you phrased the question.
Still true for simple tasks. But as AI tools get more capable - agents taking multi-step actions, tools pulling external data, workflows spanning multiple systems - the bottleneck has shifted. The model isn't the problem. The context is.
Context engineering is what happens when you take that seriously.
What is context engineering?
LangChain's definition: building dynamic systems to provide the right information and tools in the right format so that an LLM can plausibly accomplish the task.
Shopify CEO Tobi Lütke put it more plainly: "the art of providing all the context for the task to be plausibly solvable by the LLM."
Same idea. Context engineering isn't about writing better prompts - it's about designing what the model sees. Instructions, background information, user data, tool outputs, conversation history - everything that ends up in the context window before the model responds.
Gartner declared in July 2025 that "context engineering is in, and prompt engineering is out." Predicted it would appear in 80% of AI tools by 2028. The timeline might slip. The shift is real.
How is it different from prompt engineering?
Prompt engineering is about the input - how you phrase a question to get a better output. Useful. But it operates on a single turn.
Context engineering operates on the system. What does the model need to know before the conversation starts? What gets loaded automatically? What gets retrieved dynamically? What gets filtered out?
The difference matters most in complex applications:
- Prompt engineering: "Write this better if I ask more specifically."
- Context engineering: "Make sure the model always has the user's role, current project, and relevant data before it responds to anything."
One optimizes a request. The other designs the information environment the model works in.
What goes into context engineering?
Context is assembled from multiple sources. A well-engineered context typically includes:
Instructions and system context - how the model should behave, what constraints apply. Set by the tool's developers.
Retrieval-augmented data - relevant documents, records, or search results pulled in dynamically. The model doesn't need to know everything; it needs the right things at the right moment.
Tool outputs - results from function calls, API responses, code execution. Agents use these to build up context across multi-step tasks.
Conversation history - what's been said in the current session. Gets pruned when sessions run long.
User context - who the person is, what they're working on, what they care about. The layer most AI tools handle worst. Also the one with the most room for improvement.
Why does it matter now?
AI models have gotten dramatically better. The gap between what they can do and what they actually do in production is rarely about model capability anymore.
It's about context.
Ask a model "what should I prioritize this week?" without knowing who you are, what you're working on, or what constraints you're under - it gives generic advice. Same model with accurate context: specific, actionable answers.
Enterprise AI pilots have failed at high rates not because the models were wrong, but because context was missing. The model didn't know what "top customer" meant in the company's specific context. It didn't know which data sources were authoritative. It didn't know the user's actual constraints.
Context engineering is how you close that gap.
Where does personal context fit in?
Of all the inputs in a well-engineered context, personal context is the one users can directly control - and the one most often left empty.
System context is set by developers. Retrieved data is automated. Conversation history builds itself. Personal context - who you are, your expertise, your current priorities, your working style - has to come from somewhere.
Most users leave this to chance. They re-explain themselves at session start, or rely on the model to infer from conversation. Both are lossy. The model is always working with an incomplete picture.
Structured personal context - extracted from real sources, organized into layers, loaded into every session - is one of the highest-leverage improvements available. Not because it's complicated, but because almost nobody does it deliberately.
→ What personal context looks like: What Is Personal Context for AI?
→ How to build it: How to Build Personal Context for AI
