Learn · Prompt engineering & structured output · CCAR-F task 4.1

Explicit criteria beat vague instructions

Too many false positives? Say exactly what counts, what doesn't, and what to do when unsure. "Be careful" is not a criterion.

Key ideas

• Precise prompts say what counts, what doesn't, and what to do when unsure.
• "Be conservative" or "high confidence only" does not reliably fix precision; explicit categories do.
• Add explicit exclusions (style, theoretical risks) and a needs_human_review bucket for uncertain cases.
• Temporarily disable a category that produces most false positives so it doesn't erode trust in the rest.
• Give the reason behind rules, use XML tags, set a role, and put long documents before the question.

Worked example: Vague vs explicit, side by side

"Check that comments are accurate" invites opinion on every comment. "Flag a comment only when the behaviour it claims contradicts the code" gives Claude a test it can apply. The same goes for invoices: define an error as line total != qty x unit price (tolerance 0.01), or a VAT rate outside {0, 5, 12, 18, 28}%, and say to ignore formatting.

Exam traps

• Answering a false-positive problem with "switch to a bigger model" or "tell it to be careful".
• Treating a confidence threshold in the prompt as a precision fix. Define categories instead.
• Keeping a noisy category switched on: it makes people distrust the accurate findings too.

Official docs

• Prompting best practices
• Prompt engineering overview

Practise CCAR-F questions · Next lesson: Few-shot prompting for consistency