QUIZ
Prompting
8 questions on Prompting, from 8 lessons. Pick an answer in your head, then open it.
On the roadmap for: AI Engineering, ML Engineering.
Question 1 of 8
Where should stable instructions go in a long production prompt, and why?
- A. At the end, so the model reads them last and remembers them best
- B. At the start, before changing content, so caching works
- C. Anywhere: position makes no difference
- D. Repeated after every paragraph of the prompt
View answer
✓ B. At the start, before changing content, so caching works
A stable prefix makes caching work and keeps behaviour consistent; put the changing parts after it.
Lesson: Prompt Engineering Patterns
Question 2 of 8
You switch to a reasoning model but keep a prompt full of step-by-step instructions. What should change?
- A. Add even more detailed step-by-step instructions for it to follow exactly
- B. Give the goal and constraints; let it plan its own reasoning
- C. Ask it to think as little as possible
- D. Remove the goal so the model isn't constrained
View answer
✓ B. Give the goal and constraints; let it plan its own reasoning
Reasoning models plan internally, and rigid step lists can get in the way.
Lesson: Prompting Reasoning Models
Question 3 of 8
Why can 'please return JSON' still produce broken output, and what fixes it reliably?
- A. Asking isn't enforcing; constrained decoding blocks bad tokens
- B. JSON is too long for models; switch to XML instead
- C. Set the temperature to zero, and then the output will always parse correctly
- D. Models can't reliably write braces and quotation marks
View answer
✓ A. Asking isn't enforcing; constrained decoding blocks bad tokens
Grammar-constrained decoding masks any token that would break the schema, so the output always parses.
Question 4 of 8
A system prompt has grown to 20 rules, and the assistant keeps breaking one or another. What helps most?
- A. Write every rule in capital letters, so the model knows each one matters
- B. Remove the system prompt and rely on the model
- C. Group and prioritise rules, fix conflicts, test each one
- D. Add a 21st rule: always follow all other rules
View answer
✓ C. Group and prioritise rules, fix conflicts, test each one
Rules interact and compete for attention; clear structure plus tests show which ones fail and why.
Lesson: System Prompts with Many Rules
Question 5 of 8
What does an automatic prompt optimiser need to work well?
- A. A good eval set and metric, since it optimises what you measure
- B. Nothing: it can tell good prompts from bad ones itself
- C. No examples at all, so the optimiser isn't biased by the ones you picked
- D. Only a bigger model to rewrite the prompts
View answer
✓ A. A good eval set and metric, since it optimises what you measure
With a weak metric it overfits or games the score, just like any optimiser.
Lesson: Automatic Prompt Optimisation
Question 6 of 8
Your model doesn't know your company's current policies. Which tool fits best?
- A. Fine-tuning on the policy documents
- B. RAG, which supplies the knowledge at question time
- C. A higher temperature so it guesses more freely
- D. A longer system prompt that contains every policy in full
View answer
✓ B. RAG, which supplies the knowledge at question time
Knowledge gaps call for retrieval; behaviour and style gaps are where fine-tuning helps.
Question 7 of 8
You need the model to always answer in a specific tone and format, across thousands of cases. When does fine-tuning beat prompting?
- A. Never: prompting can always fix tone and format alone
- B. When prompts aren't consistent enough and you have examples
- C. When you have no examples of the behaviour you want
- D. When you need the model to learn new facts that keep changing week to week
View answer
✓ B. When prompts aren't consistent enough and you have examples
Fine-tuning changes behaviour; it is a poor way to add facts that change.
Question 8 of 8
A one-line prompt fix passed five spot checks but broke other cases in production. What process prevents this?
- A. A regression eval on every change, plus staged rollout
- B. More spot checks by hand before each release
- C. Never changing prompts once they seem to work
- D. Testing only the exact case that was reported to you
View answer
✓ A. A regression eval on every change, plus staged rollout
Prompts are code: changes need tests over many cases, not a handful of examples.