Prompt Design and Effective Use of Generative AI Tools
Domain 3 tests practical prompt engineering: can you turn a vague request into a specification a model can follow, and can you explain why one prompt outperforms another? Questions typically show two versions and ask which change improves reliability, or ask for the control that prevents a specific failure.
Anatomy of a strong prompt
- Task: one clear verb and deliverable.
- Context: audience, purpose, background the model cannot infer.
- Input: the source material, fenced with delimiters such as triple quotes or XML tags.
- Constraints: length, tone, reading level, what to exclude.
- Format: explicit structure - JSON keys, table columns, headings, word counts.
- Examples: one or two few-shot examples fix style and format far better than adjectives.
Techniques that earn marks
- Delimiters separate instructions from data and reduce accidental instruction-following.
- Schema with a null rule (return null if the field is absent) cuts fabrication in extraction tasks.
- Grounding: answer only from the text provided, quote the supporting sentence, otherwise reply that it is not in the source.
- Decomposition: split multi-step work into a plan and then execution, rather than one giant instruction.
- Reasoning prompts: ask for steps where judgement benefits; ask for direct output where an invented rationale would add noise.
- Iteration: change one variable at a time and score against a fixed test set.
Worked example
Weak: 'Write something about our new expenses policy.' Strong: 'You are an internal communications writer. Using only the policy text between the tags, write a 150-word email to non-managers. Plain English, no legal jargon, one clear call to action. Format: a subject line, then three short paragraphs. If a detail is not in the policy, do not invent it - say so instead.' Reasoning: the strong version names the role, locks the source, sets audience and length, prescribes structure, and adds an anti-fabrication rule with a fallback behaviour. Each element removes a specific failure mode; 'write something' removes none.
Traps that cost marks
- Adding 'be accurate' or 'this is very important' and expecting a mechanism.
- Believing politeness or threats improve correctness.
- Stacking contradictory rules in a long prompt; models weight some instructions more than others.
- Treating a system prompt as a security boundary - it is a preference, not enforcement.
- Not re-running the test set after a model upgrade.