Page 83 - CyberSecurityEssentialsEnglish

use more deterministic settings if the tool allows. Prompt engineering is not a test of the
text alone; it is a test of how the prompt interacts with the model, its settings, and the data.

↻ Try it yourself: a single-change experiment
Choose a weak prompt. Record the first output, identify the single largest defect, then
change one element only. Evaluate the second version against the same list. Write one
line: what did you learn about the task or about the model?

FIGURE 2-6: DIAGNOSE THE CAUSE BEFORE YOU REWRITE THE PROMPT

What to notice: Rewriting is one remedy among five. Naming the cause first prevents the
same edit being applied to every kind of failure.

Source: Prepared by the author.

● Remember: Section summary
• Key idea: Improvement comes from diagnosis, so change the one factor that explains
the defect.
• Practical decision: Name the largest defect, change a single element, then compare the

two versions against the same list.
• Mistake to avoid: Rewriting the whole prompt after every output, which hides what
actually caused the change.

2.9 Prompting Reasoning Models and Structured Outputs
◆ Key idea: a reasoning model needs a clear end point, not a longer
instruction.

Models built for reasoning already work through steps internally, so repeating an

82