Page 94 - CyberSecurityEssentialsEnglish

Begin with five recurring tasks that consume time: summarizing a particular kind of report,
preparing a follow-up email, extracting tasks from minutes, or turning notes into a plan. Do
not collect a hundred templates from the internet. A template you have neither used nor

tested does not represent productive knowledge.

Make the template parameterized, with fields such as [audience], [length], [material], and
[tone]. Parameters prevent a client name or a date from surviving out of a previous use, and

they make clear what must be changed. Store a successful example and a known failure
alongside it, not the text alone.

Use versioning for the templates that matter. Record the version number, the change, the
reason for the change, and the result on the evaluation cases. You do not need a complex

software system; a simple table is enough for personal use. The aim is to know what
improved and why.

Test again when the model, the tool, or the type of input changes. A prompt that suited one
version may become redundant or less effective with a new one. Treat the template as a

living asset rather than a fixed text. In organizations this becomes the management of
prompts, evaluations, and approvals within what is often called GenAIOps, meaning the
operational practices that govern generative AI systems in production. That subject lies

beyond the scope of this chapter.

Portability Between Models and Tools
Do not assume that a template that succeeded with one model will succeed in the same way
with another. Models differ in instruction following, in the use of examples, in the size of the

context window, in tool support, and in how they handle images and structured outputs.
Keep the objective and the criteria fixed, then test the basic version on each tool instead of
carrying over dozens of tricks that suited an older model.

When you migrate, begin by removing redundant instructions. A newer model may be better

at understanding direct natural language, and some older phrases become a source of
confusion. Then add back only what the evaluation results require. This process is sometimes

called prompt migration, and it is a comparison of behavior rather than a literal copy.
Record the name of the model or the service and the settings that matter alongside the test
result. The template alone does not explain the output; the result is affected by the model,
its version, its settings, the tools, and the sources. When one of these changes, retest the

critical cases before relying on it.

Template Privacy and Safe Reuse
Clean the template of the names of people and clients and of confidential numbers before
saving it. Replace them with parameters, and describe the type of data required instead of

93