Page 97 - CyberSecurityEssentialsEnglish

2.14 Practical Lab: From a Weak Request to a Tested Professional Prompt

↻ Try it yourself: the output of this lab
By the end of the activity you will have: an initial request, an expanded PACFT prompt, an
acceptance list, at least five evaluation cases, two versions to compare, and a saved
template with its limits stated.

Step One: Choose a Low-Risk Task
Choose a task you repeat and whose result you can evaluate: summarizing a report,
preparing an email, extracting fields, building questions, or organizing a meeting. Avoid

sensitive data and high-impact decisions in the first experiment. Write the request the way
you would normally write it, in one line, and do not improve it yet.

Step Two: Write the Purpose and the Success Criterion Before the Prompt
• What will happen after the output is used?
• Who will read it or act on it?

• Which three to seven items make the result acceptable?
• Which error cannot be accepted?

Step Three: Apply PACFT, Then Add the Complementary Elements
Write the purpose, the audience, the context, the format, and the tone. After that, add the

specific task, the source material, the constraints, and the success criteria. If there are gaps
that matter, ask for clarifying questions and stop before producing anything.

Step Four: Build a Small Evaluation Set
Gather at least five cases: an ordinary case, a long case, an incomplete case, an edge case,
and a case containing Arabic terminology or a complex layout. Write the standard for the

result before running the tool. Do not change the success criterion after seeing the output,
so that you do not bias yourself toward the result you got.

Step Five: Run and Diagnose
Run the initial request and then the structured prompt on the same cases. Evaluate each

output against the list. Identify the largest defect in understanding, source, format, tone, or
stability. Record the time spent writing and reviewing, because the goal is the time to a valid
result, not the speed of text generation.

Step Six: Improve One Factor

Add an example, a criterion, or a policy for missing values, or remove a conflicting constraint.
Retest and compare. If the failure persists, use the diagnostic tree: do we need a current
source, a different model, a calculation, a schema, or a workflow?

96