Page 69 - CyberSecurityEssentialsEnglish

What to notice: Prompting resolves controllable ambiguity. It does not create current
knowledge, and it does not turn a generative tool into a calculation engine or a reliable
execution system.

Source: Prepared by the author based on published guidance on the limits of prompt
optimization [4].

TABLE 2-4: WHEN IMPROVING THE PROMPT IS NOT ENOUGH

Problem The sign A possible remedy

The task is The output is generic, or changes Sharpen the objective, the context,
ambiguous with every attempt. and the criteria.

Attach a source, use search, or use
The knowledge is
The system guesses, or answers retrieval-augmented generation,
unavailable or too
with outdated information. which Chapter Three covers in
recent
detail.

It fails on the language, the
The model is Change the model, or use a
medium, or the complexity
unsuitable specialized tool.
despite a fair prompt.

The task needs Numerical or logical errors appear Use a calculator, code, or an

exact calculation in results that can be executed. analysis tool.

The structure must Field names change, or elements Use schema validation or structured
be strict go missing. outputs.

Performance is Build an evaluation set, then
One example succeeds while
unstable across compare tools or consider
other cases fail.
cases specialized tuning.

The work is multi- State is lost between rounds and Turn the task into a workflow, as
step and repeated errors recur. Chapter Five describes.

● Remember: an escalation rule prepared by the author
If an output needs correcting for the same reason more than twice, turn the repeated

instruction into a constraint, a criterion, or an example inside the template. If it still fails
after that, examine the source, the model, and the tool instead of continuing to patch the
prompt.

2.5 The PACFT Model: A Quick Recall Tool by the Author

◆ Key idea: PACFT is a memory aid, not a standard.

It gathers five questions that recur in everyday prompting: purpose, audience, context,
format, and tone. It does not replace the source material, the constraints, or the
acceptance criteria, and it cannot make a badly defined task a good one.

68