Context engineering for teams that don't write code

“Context engineering” has replaced “prompt engineering” as the phrase of the year among people who build with AI. The idea is simple: the answer depends less on how you ask and more on what the model has in front of it when it answers. Which documents, which instructions, which tools, which facts.

Most of what’s written about it is for developers. Token budgets, retrieval pipelines, memory compaction. If you run a sales team or a support desk, you can skip all of that. The part that matters to you fits on one page.

The difference to prompt engineering

A prompt is the question. Context is everything the AI knows when it reads the question.

You can write a perfect prompt about your refund policy and still get a wrong answer, because the AI has never seen your refund policy. It fills the gap with what’s typical. Better prompts don’t fix missing context. They just make the wrong answer sound more polished.

What context means in a company

For a business team, the context that goes wrong is almost always the same kind: the rules you work by. What we offer and what we don’t. Which discounts need approval. How long delivery takes. Which words legal doesn’t want in an email.

That kind of knowledge is hard on AI. It sits in people’s heads, in a wiki and in old emails. It changes without anyone announcing it. And people disagree about it more often than they’d admit.

Three habits that do most of the work

Write rules down as rules. Not as a 40-page handbook, but as short statements with a reason. “Custom orders can’t be returned, because we can’t resell them.” The AI can use a sentence like that. It struggles with chapter 7.3 of a PDF from 2023.

Give every rule an owner. Someone decides what applies. When sales and support disagree about the return window, that’s a conversation between two people, and then one person updates the rule. Without an owner, the AI ends up as the referee, and it picks whichever version it found first.

Give every AI the same source. If each person pastes their own version into their own tool, you’ve engineered five different contexts. Connect the tools to one place instead. Today that usually means MCP, which Claude, ChatGPT and Copilot support in their business plans.

What you can leave to the tools

You don’t need to decide how many tokens fit, how documents get split up or which retrieval method to use. Claude, ChatGPT and the tools built on them handle that, and they get better at it every few months.

What they can’t do is decide what’s true in your company. That stays with you. It’s also the part that keeps its value when the next model comes out, which is why we think it deserves more attention than the next prompt trick.

A small test

Pick five questions your customers or colleagues ask every week. For each one, write down where the correct answer is written today and who would have to approve a change. If you can’t name a place or a person for three of the five, your AI is working without context on those questions, however good it is.

Where to start

Start with the five questions from the test. Write each rule once, with a reason and an owner, and make sure the AI tools your team uses read that version.

For a handful of rules a shared document is enough. Once you have a few hundred, several owners and more than one AI tool, you need something that tracks versions, flags contradictions and serves the current state to every tool. That’s what a company brain is for, and it’s what we build with CtxCore Brain.