The context on context.

You hear about context engineering, context windows and giving a tool more context, usually without anyone stopping to say what the word means. A good place to start is a simple question: when you ask an assistant about a company, where does the answer come from?

1. Training and context

Every answer draws on two sources. The first is training: before a model is released, it reads an enormous amount of text and learns patterns from it, and that learning is then frozen at a cutoff date.

Anything that happened after the cutoff is unknown to it. If your prices changed last month, the model’s training has no way of reflecting that.

The second source is context, which means everything in front of the model at the moment you ask: your question, the conversation so far, any instructions the tool was given, and any files or search results it has been handed. Context changes with every request, which is what makes it so useful.

2. The desk

The easiest way to picture the difference is a desk. Training is what the model remembers: broad, general, a little hazy and fixed as of a date. Context is what is on the desk when it sits down to work, and that can be specific, current and different every time.

When the two disagree, a model tends to lean on the page in front of it rather than on its memory. It does not do this every time, but it does it often enough that the same assistant can give a sharp answer with the right document attached and a vague one without it.

That changes the question worth asking. What the model remembers about you is hard to change, so the more useful question is what is on the desk when it works on you.

3. Usually, whatever turned up

For a lot of companies, the honest answer is whatever the tool could find or whatever someone happened to paste in. That might be the current pricing page along with the archived one, last year’s sales deck, a review listing, a forum thread and a comparison post written by somebody else.

None of those pages is on the desk to cause trouble. They are old, incomplete or written by someone who is not you, and the model does its best with them. Its best is usually a confident blend of all of them.

4. More material makes it harder

Once people understand this, the natural instinct is to put everything on the desk: every deck, every document, every style guide and every transcript.

In practice that tends to make answers worse. The desk has a size limit, called the context window, and even well within that limit, more material gives the model more contradictions to resolve. A handful of documents that agree will usually do better than a pile that overlaps, because the model is no longer choosing between versions of the same fact.

The goal, then, is the right context for the task, kept as small as it can be while still being complete.

5. What good context looks like

For a company, good context has five properties:

• Current and dated, so nothing stale sits next to something new.
• Sourced, so every fact can be traced back to where it came from.
• Approved, so what the model repeats is something a person with the authority to agree to it has agreed to.
• Consistent, so every tool reads the same version.
• Compact, so it fits on the desk without burying the thing that matters.

Put those together and you have a description of a record: one place, written for machines to read, where each fact appears once.

6. Two kinds of desk

There is an honest limit to all of this. You control the desk for the tools your company runs: your team’s assistants, your agents and your workspaces. Those can read your record every time they are used.

You do not control the desk of a stranger’s assistant. What you can do is make sure the best material it can find about you sits on your own domain, current and plainly written. That will not decide what it says, but it does improve what it is likely to be reading when it says it.

7. So, the context on context

Every company has access to the same models, and the training behind them is shared by everyone. The desk is where companies can actually differ, and that is what context means in practice. It is also the part you control.

Start with the diagnostic. It is free, it is one pass, and you keep what it finds whether or not anything follows it.