Prompt engineering

Prompt engineering usually brings to mind the question a user types. In an enterprise product the decisive layer is the system instruction: the frame the model sees in every conversation and the user cannot change. Scope, tone, boundaries, and the obligation to cite are defined there.

In KobiGPT every department template ships with its own system instruction. The accounting assistant works in a VAT and tax context, the HR assistant in a labour-law context, the legal assistant in a data-protection context. That removes the need for users to restate the context every time.

A good system instruction states not only what to do but what not to do. In KobiGPT templates this appears as the behaviour of stating that information is unavailable rather than guessing when nothing is found.

That choice lowers the answer rate but raises trust. In a corporate knowledge base, saying "I do not know" beats inventing — because an invented procedural answer can turn into an operational error.

Some problems cannot be prompted away. If a fact is not in your document archive no instruction can produce it; if retrieval fetches the wrong passage, a prompt will not fix it. In those cases the fix belongs on the document or indexing side.

A practical approach: when answer quality is poor, look at the sources first. Are the document passages shown alongside the answer the right ones? If not, the problem is retrieval, not the prompt. If they are right and the answer is still weak, the problem is the document itself.

A useful way to evaluate prompt-engineering is to follow one real question from the source document to the final answer. Record which file was selected, what context reached the model, and what a reviewer would need to verify. This turns a definition into an operational check and makes the result comparable across teams.

The same check should include ownership and change management. Decide who updates the relevant documents, how an outdated result is reported, and which access boundary applies. KobiGPT can provide the assistant and the cited document context, but the organisation still owns the source material, permissions, and the decision made from the answer.