Page 103 - CyberSecurityEssentialsEnglish

Practical Exercises

↻ Exercise1: Realistic improvement
Goal: Turn a request you already send into a prompt that states what a good result would
be.
Required: One short request you use at work. Allow about 25 minutes.
Steps: Write the purpose in one sentence. Attach or describe the source material and say
whether outside knowledge is permitted. Describe the audience by their knowledge,

their time, and the decision expected of them. Fix the format and the length. Then list
three to five acceptance criteria.
Output: The original request and the rewritten prompt side by side, with the acceptance
list beneath them.
Assessment criterion: Another person could judge the output against your acceptance list
without asking what you meant.

↻ Exercise 2: PACFT and its limits

Goal: Use a framework as a recall tool without treating it as a standard.
Required: One task from your work. Allow about 20 minutes.
Steps: Write the prompt using purpose, audience, context, format, and tone. Run it once
and keep the output. Then read the prompt against the task and list what the five letters
did not capture, such as the source material, the ranking of conflicting instructions, and
the acceptance criteria.

Output: A prompt built on the five elements, and a list of what they left out.
Assessment criterion: The list is specific to the task. Concluding that the framework
covers everything means the task was not tested against it.

↻ Exercise 3: One example and a few examples
Goal: Measure what examples add instead of assuming it.

Required: A classification task with at least twelve items whose correct labels you already
know. Allow about 30 minutes.
Steps: Run the task with no example, then with one example, then with three examples
chosen to differ from each other. Keep everything else identical. Record the accuracy and
the kind of error at each stage.
Output: A three-row comparison of accuracy and error type.
Assessment criterion: The three runs differ only in the examples, and the errors are
described rather than only counted.

↻ Exercise 4: Settings experiment
Goal: See that generation settings change variety, not knowledge.
Required: A tool that exposes temperature or nucleus sampling, and one prompt with a
factual answer. Allow about 20 minutes.
Steps: Run the prompt twice at a low setting and twice at a high setting. Record what

102