Context engineering for a small business

Here is the complaint it answers: the AI doesn’t understand the context of what I’m doing. The answers are fluent, general, and wrong in the specific ways your business is specific. So people re-explain — the client, the pricing, the history — every single session, from scratch.

Context engineering fixes that by moving the explanation out of your typing and into files. You write down, once, the things you keep saying: how the work gets done here, what was already decided, which clients are special cases. The tool reads those files at the start of the work. From then on it answers like someone who has been at the company for a while, not someone who walked in this morning.

The observation this page rests on: most AI disappointment in a small business is missing written context, not a weak model. The models are already capable. What they are missing is what only you can write.

Five, to start. Each is a page or two of plain text, written by someone who knows the business, updated when reality changes.

Notice what is not on the list: your whole document archive. Dumping everything into an AI tool buries the five files that matter under a thousand that don’t. The craft is selection — writing down the load-bearing knowledge and leaving the rest in the filing cabinet where it belongs.

Prompt engineering improves one request. Context engineering improves every request, because it changes what the tool knows before any request arrives. A well-crafted prompt over empty context produces a confident, articulate guess.

There is a second difference people feel before they can name it. Without your decisions written down, an AI tool is a yes-man: it agrees with everything, because it has nothing of yours to check a new idea against. Give it your decisions and exceptions and it can finally push back — poke holes in a plan that contradicts a call you made in March, or flag that the discount you are about to offer breaks your own pricing rule. The file is what gives it permission, and material, to call you out.

Prompts are still worth writing well. They just cannot carry knowledge they were never given. Why sessions lose that knowledge in the first place is covered in why AI forgets.

Start with one file, not five. Pick the question your team answers most often — for most businesses it is some version of how do we handle X — and write the one page that answers it. That is usually the how-we-work file.

Then run the comparison. Ask your AI tool a real question from this week’s work without the file, then again with the file loaded. The difference is the argument, and it is usually visible on the first try. Add one file a week after that: decisions next, then exceptions. Five weeks in, the tool has stopped guessing about your business.

Keep every file in plain text you own, not inside any one product’s memory — that keeps the work portable across models and vendors, which is the same principle behind a personal AI brain. And if the honest problem is that nobody can name what is missing, that mapping is a bounded job with a fixed price: the Working Intelligence Audit.