Page 89 - CyberSecurityEssentialsEnglish

TABLE 2-14: REWRITING QUESTIONS TO REDUCE SYCOPHANCY

Loaded wording More neutral wording

Compare A and B against cost, time, and risk, then state what
Isn’t option A clearly better?
the data does not allow you to settle.

Identify the strongest reasons it might succeed and the
Confirm that my plan will
strongest reasons it might fail, and the small test that would
succeed.
distinguish between them.

Give me a 95 percent State the assumptions, the evidence, the part most exposed to
confidence level. error, and what could change the result.

Defend my view in front of Build the best case for the view, then the best independent
the team. objection, and propose data that would settle it.

● Remember: Section summary
• Key idea: Useful review needs a criterion, a source, or a test, and sycophancy can make
agreement look like independent analysis.
• Practical decision: Separate critique from rewriting, and ask for the evidence, the
assumptions, and what could change the result.
• Mistake to avoid: Treating a system’s review of itself, or its stated confidence number,
as a calibrated certificate of correctness.

Video: Prompting 101
A practical workshop showing how context,
instructions, examples, and output format change
the quality of a result, with prompts improved step

by step.
Duration: about 24 minutes | Level: Beginner to
Intermediate
Learning objective: Follow a full cycle of building

and refining a prompt on a real task.
Source: Anthropic [20].
https://www.youtube.com/watch?v=ysPbXH0LpIE
Verified: June 6, 2026

2.11 Managing Long Tasks and Large Documents
◆ Key idea: a large context window is not a perfect memory.

A model may accept a long file, but it does not use every part of it equally. Information
can be lost in the middle, versions can be mixed, and conflicting instructions can
accumulate. We therefore need a map of the material, purposeful chunking, uniform
extraction, a working brief, and staged verification.

88