Important, but not the part that makes a business agent good.
Prompt engineering is the careful formulation of instructions for a language model to get useful answers: role, context, desired format and boundaries. With a business agent this is one part of the setup; the sources he is allowed to consult and the boundaries he is not allowed to cross are more important.
IN COMMON LANGUAGE
Which contains good instruction
A clear role, the context in which the answer must be given, the desired format, and explicitly what must not happen.
The latter is most often forgotten. A model that is not allowed to explicitly say that it does not know something prefers to come up with something plausible.
WHY IT MATTERS
Why it is not the core for us
A perfectly worded instruction about wrong or missing sources still gives the wrong answer. Organizing sources yields more than formulating.
The order that works: first make the right information available, then determine the boundaries, and only then fine-tune the formulation.
- Role and context — from what position the answer is given.
- Explicitly state what is not allowed — otherwise the model is guessing.
- Allowing not knowing — the most important instruction.
- Desired format — short, with source, in Dutch.
- Sources carry more weight — the wording comes next.
FREQUENTLY ASKED QUESTIONS
More about prompt engineering
Do I have to learn this?
It certainly helps for daily use of an assistant: providing context and saying what you want produces better answers. We do that work for a business agent, and the profit mainly lies elsewhere.
Does this change with new models?
The details yes, the principle not. Being clear about role, context and boundaries continues to work, even as models require less guidance.
Is a long instruction better?
Not automatically. Too many instructions compete with each other and make behavior unpredictable. Short and explicit works better than long and exhaustive.
CONTINUE READING
Of course, continue reading about this subject.
AI that one usable answers?
We equip agents with the right resources, boundaries and escalation: that's where the real work lies.