What Your AI Assistant Needs to See Before It Can Help
An introduction to context engineering: what background information actually changes an AI system’s output, and what to leave out.
Who it is for
Owners and operators setting up an AI assistant to work over their own material.
After reading it
You will be able to assemble scoped, current context for a task and recognize the signs that an assistant is missing it.
There is a difference between phrasing an instruction well and giving a model the right environment to work in. The first is wording. The second is context engineering, which means designing the full set of information a model works from at the moment it acts: the background, the relevant history, the source documents, and the constraints. A well-worded question over the wrong context still produces the wrong answer.
Context is the environment, not the sentence
Think of context as everything the model can "see" while it answers. If your assistant does not know your current prices, your service area, or last week's decision, no amount of clever phrasing will make its answer correct. The fix is not a better sentence. It is a better working environment: put the price sheet, the relevant note, and the rule in front of it.
What belongs in context, and what does not
The instinct to paste in everything you have is understandable and usually wrong. More text is not more help. It buries the signal and pushes the model toward the wrong details.
- Include: the specific records the task touches, the constraints that must hold, and the recent history that changes the answer.
- Leave out: unrelated documents, old versions, and background that does not bear on this task.
Dumping an entire fifty-page document in often produces worse results than a scoped two-paragraph excerpt, because the model has to find the relevant part itself and can land on the wrong one.
A checklist before you ask
Before you send a task to an AI assistant, run a short mental check:
- What does the model need to know that it cannot know on its own?
- Which one or two documents actually bear on this?
- What constraint must not be broken (a policy, a budget, a tone)?
- What is the most recent fact that changes the answer?
If you can answer those in a few lines, you have your context. Paste those lines, not the whole archive.
Signs your assistant is missing context
You can usually tell when context is thin. The answers come back generic. The model makes an assumption that would have been obvious to anyone who knew your situation. It asks the same question twice, or it confidently states something that was true last quarter but not today. Each of those is a signal to add the missing piece rather than to reword the request.
A short example
Ask "what should I charge for a consult" with no context and you get a generic range. Give the model your service list, your market, and the note that you do not publish prices, and ask it to draft how you would answer a prospect who asks about cost, and you get something that sounds like your business, because it was working from your business.
Where this leaves you
Good phrasing helps. The right context decides. When an AI assistant answers well, it is usually because someone gave it the right things to look at, scoped and current, and left the rest out.