What context does AI need to do useful work?

Why a capable assistant can still give a generic answer

Imagine asking a new colleague to prepare a management update. They can write well, but they have not attended the meetings, do not know which issue has already been resolved and have never seen an update your manager found useful. A polished first draft could still be wide of the mark. Much everyday AI disappointment has the same explanation.

A general-purpose AI system can work across many subjects. That does not establish its knowledge of your organisation or the particular decision in front of you. When someone asks it to "make this more strategic", the missing information may be what the decision-maker needs to decide and what evidence would change that decision.

Context engineering is the term often used for arranging the information and instructions an AI system receives. For a business user, the useful question is straightforward: what does this assistant need to know, where will it get that information, and how will I know whether it used the right material?

Four different ways information reaches the assistant

These are practical distinctions, not a promise that every product exposes four separate controls. The features available depend on the application, account and organisational settings. Identify the actual setup before assuming that an answer used company files or remembered an earlier decision.

Separate standing instructions from changing facts

An instruction such as "write for colleagues who have not attended the meeting" can remain useful across many updates. The date of the next meeting belongs with the current task. Combining both in one permanent instruction creates a maintenance problem: next month, an obsolete date may still look like a rule.

Claude Projects, for example, provide project instructions and a knowledge area that can be used across chats in that project. Those are useful places for an agreed working approach and relevant reference material. They still need an owner. A shared project that contains last year's policy requires the same attention as a shared folder containing it.

For a recurring report, explain the audience and reporting conventions separately from the period being reported. State which document governs if the supporting material disagrees. Where the disagreement is unresolved, the assistant should expose it for review. Asking it to choose the most plausible version can conceal the very issue the reader needs to see.

A large library does not mean every document was read

Anthropic describes retrieval in Projects as a way to select relevant material from a larger knowledge collection instead of placing the entire collection into the working context. Microsoft describes grounding as adding relevant material to a Copilot request before the model generates its response. Both help explain why storing a document and using it in a particular answer are different events.

If an answer omits a critical exception, establish whether the source was found and whether the relevant passage was used. A more elaborate writing instruction will not resolve every retrieval problem. Sometimes the useful change is to identify the governing document directly or narrow the question to the material that matters.

Be careful with requests for exhaustive coverage. "What issues appear in these documents?" and "Have you checked every applicable clause?" carry very different expectations. A fluent overview is weak evidence of completeness. For consequential work, define the scope that a person will verify.

Memory helps continuity, but needs review

Claude offers memory and past-chat search, with behaviour and controls that vary by account. Its documentation distinguishes project memory from other conversations. This can reduce the amount of background someone has to repeat, but an assistant remembering something does not establish that it remains true.

A remembered reporting preference is relatively easy to correct. A remembered approval, deadline or policy interpretation can affect someone else's decision. Put consequential facts in an identifiable current source and review them when they are used. Keep memory for continuity without treating it as an approval record.

Decide what context is worth maintaining

Start with a recurring job where people repeatedly explain the same background. Agree what a new person would need to understand, which documents they should trust and who can resolve exceptions. This is often a manageable improvement within the tools already approved for the team.

Too much undifferentiated material can make the task harder to inspect. Prefer a clearly described set of relevant sources, with enough information to locate the originals. Protect sensitive information using the organisation's approved systems and access arrangements.

Judge the improvement by the work: fewer missing facts, less repeated explanation and a result that the intended reader can use. A longer prompt is only useful if it helps achieve that. Product documentation below was checked on 3 October 2026; verify current account features before standardising a team method.