What Is Context Engineering? A Plain-English Guide (2026)

For about two years, the advice was to get better at prompts. Learn the magic phrasing. Add "think step by step." Tell the model it's an expert.

Then the people building serious AI systems started saying the phrasing was the small part. What decides the quality of an answer is everything the model is reading when it writes it. They gave that work a name, context engineering, and by 2026 it's the term you'll hear in every conversation about AI agents.

Most of what's written about it is for developers. Here's what it means in plain English, and what the version for one person looks like.

What is context engineering?

Context engineering is the practice of deciding what information an AI model has in front of it when it answers, and keeping that information right.

A language model doesn't know anything about your situation except what's loaded into its context window at that moment: the instructions, the conversation so far, any files, search results, tool outputs, and saved memory. Context engineering is the job of choosing that material on purpose instead of leaving it to chance.

The term went mainstream in June 2025. Shopify CEO Tobi Lütke wrote on X: "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."

Andrej Karpathy backed it a few days later, describing context engineering as "the delicate art and science of filling the context window with just the right information for the next step." Simon Willison, who has written about AI tools for years, noted the shift that same week.

Then Anthropic made it formal. Its September 2025 guide, Effective context engineering for AI agents, defines it as "the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference."

Three sources, one idea. The words you type are a fraction of what the model reads. The rest is the real lever.

Context engineering vs. prompt engineering

	Prompt engineering	Context engineering
Focus	How one request is worded	Everything the model reads alongside it
Typical question	"How should I phrase this?"	"What does the model need to know, and what should it not see?"
Scope	A single message	Instructions, history, files, memory, tool results
Time horizon	This answer	Every answer, across a whole session or system
Common failure	Vague or ambiguous wording	Missing, stale, or cluttered background
Who does it	Anyone typing into a chat	Developers building agents, and anyone who sets up an assistant on purpose

Prompting didn't go away. A clear request still matters. But a perfectly worded question on top of the wrong background gets you a perfectly worded wrong answer.

Why did context engineering take off?

Because of agents.

A chatbot answers one question. An agent works through a task over many steps, calling tools, reading results, and deciding what to do next. Every step adds material to the context window, and the window has a limit. Anthropic's guide puts it bluntly: context "must be treated as a finite resource with diminishing marginal returns."

That second half is the part people miss. More isn't automatically better. Past a point, extra material buries the parts that matter, and the model's attention spreads thin. Anthropic's guiding principle is "finding the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome."

So the real skill is editing.

What actually goes into an AI's context?

Whether you're building an agent or typing into ChatGPT, the same handful of things end up in the window:

• Instructions: the system prompt, custom instructions, or project settings that frame every answer.
• Conversation history: what's been said in this session, until it grows too long and the oldest turns fall out.
• Files and retrieved documents: anything uploaded, attached, or pulled in by search.
• Tool results: output from a web search, a calculator, a database, or another app.
• Memory: short notes the app saved about you and quietly re-inserts later.

If the difference between the window and memory is fuzzy, this breakdown of context window vs. memory walks through it. The short version: memory is just another way of putting text into the same window.

What does context engineering mean if you're not a developer?

You're already doing it. Every time you paste a paragraph of background before asking for help, you're engineering context by hand. Usually in a hurry, usually from scratch, and usually leaving out the part that would have changed the answer.

The individual version comes down to three decisions:

1. What should an AI always know about me? Your role, your values, how you write, what you're working on this quarter, the people who show up in your work.
2. What does it need for this task only? The draft, the data, the client brief. Load it when you need it, not before.
3. What should it never see? Old priorities, finished projects, detail that wouldn't change the answer. Leaving things out is half the job.

Picture a hypothetical school principal who uses AI to draft parent newsletters, board updates, and staff emails. Without any setup, every draft sounds like a press release. With a one-page document covering her school's values, her plain and warm writing style, this year's two big initiatives, and a note never to name individual students, the same model writes something she'd actually send. Nothing about the model changed. The context did.

That permanent layer is what RUMO calls context anchors: short documents you own, each covering one part of who you are and what you're doing.

How do you do context engineering for yourself?

Start by noticing what you keep re-explaining. That list is your raw material.

Sort it into the always-true and the just-this-once. Write the always-true part down once, in a plain document that lives outside any single app, so it survives the day you switch tools. If a blank page stops you, that's normal, and pulling it out through guided questions has a name: context mining. The free Personal Constitution builder runs you through the values layer in about half an hour.

Then cut. If a line wouldn't change how an AI answers you, it's noise. Load what's left at the start of a session and add only the task material on top.

And keep it current. Developers worry about stale data in a pipeline. You have the same problem with stale priorities. Context about the present expires, and a few minutes every few weeks keeps the AI from confidently working off last spring's version of you.

Is context engineering the same as HubSpot Growth Context?

Same discipline, different scale. HubSpot built its whole fall 2026 release around Growth Context, the business, team, and customer knowledge its AI runs on, with a score showing where the gaps are. That's context engineering for a company.

The personal version covers one human being. And the underlying distinction is the one between data and context: the record of what happened versus what it means for you right now.

The short version

Context engineering is deciding what an AI knows when it answers. Developers do it with retrieval pipelines and token budgets. You can do it with a single document.

The model is going to read something before it replies. The only question is whether you chose it.