Page 92 - CyberSecurityEssentialsEnglish

+ For the advanced reader: the order of material may affect performance
Models differ, but it is useful to test where the question, the documents, and the
examples sit rather than assuming a single order. Some guidance places long documents
first and the question at the end, with a request to extract the relevant passages before
answering. This is a practice that needs testing on your model and your task; it is not a
fixed general rule.

● Remember: Section summary
• Key idea: A large window allows more input; it does not guarantee equal attention to all
of it.

• Practical decision: Map the material, chunk by meaning, extract with one fixed
template, and verify at every stage.
• Mistake to avoid: Sending an entire archive and expecting a final conclusion with no
evidence trail.

2.12 Common Mistakes and the Limits of Prompting

◆ Key idea: not every failure is a failure of the prompt.
Some failures come from the design of the task, some from expecting knowledge the
system does not hold, and some from asking a single request to do the work of a whole
process.

Prompting errors are not only errors of language. Some come from the design of the task,
some from unrealistic expectations, and some from confusing the improvement of a text

with the construction of a reliable system. The following list helps with rapid diagnosis, but
privacy, verification, security, and workflow are treated in far greater depth in later chapters.

91