Page 66 - CyberSecurityEssentialsEnglish

The Order of Instructions and a Policy for Ambiguity

A prompt does not operate in a vacuum. In many applications there are instructions above
the user message, such as system instructions, developer instructions, or tool policies. The
conversation may also carry earlier context. This is why a system may refuse part of a

request, preserve a particular format, or follow a rule that does not appear in the current
message. Understanding this hierarchy prevents every difference from being read as a
weakness in the user’s wording.

Within your own prompt, order the instructions from most to least important, and do not

leave a conflict without a policy. You can write: “The first priority is preserving the facts, then
clarity, then brevity. If the word limit conflicts with including the facts, exceed the limit by
no more than 10 percent and say so.” In this way the system is not forced to choose one

condition in silence.

Add an uncertainty policy when the task demands precision: “If the request admits two
readings that would change the result, ask before continuing. If the information is not in the

source, write not stated rather than guessing.” This policy does not make the system
infallible, but it makes the required behavior explicit when information is missing or
contradictory.

You can also define the boundaries of scope: what falls inside the task and what falls outside

it. For instance: “Review the clarity and structure of the text, and do not assess the accuracy
of the medical claims.” Separating scope prevents the system from offering an assessment
that looks comprehensive when it was given neither the evidence nor the right tool. In high-

impact tasks, the point of handover to an expert or an external source should be stated
explicitly.

FIGURE 2-3: THE ANATOMY OF A PROFESSIONAL PROMPT

65