Aptitude · Prompt engineering
Prompt engineering, made easy
💡 The idea
In a hiring test, your prompt is run on hidden inputs and the reply must match exactly. A good prompt leaves the AI nothing to guess: what the input looks like, what to do with tricky cases, and exactly what to output.
Task→Input→Rules for tricky cases→Exact output + ONLY→Example→{input} last
🧭 The method: 3 steps
- Say what the input looks like and what to do with it.
- Add a rule for every tricky case: repeats, ties, capitals, empty input, rounding, boundaries.
- Give the exact output format with “Reply with ONLY…”, one tiny example, and put {input} last.
🎬 Every kind of question, in pictures
Pick a kind. Guess the answer first, then watch the picture work it out. Tap any step to see it in the picture.
Loading the pictures…
✏️ Worked examples
Example 1. “Who won the vote? {input}” fails hidden tests. Why?
- No output format: the AI writes a sentence
- No rule for a name on several lines (add them up)
- No rule for ties
- Better: “Each line is
: . Add up repeated names; ties → first in the list. Reply with ONLY , (e.g. Ana,12). Votes: {input}”
Example 2. The AI replied “```json [...] ```” but the test failed. Fix?
- The answer itself was right
- The ``` code block is extra text
- Add: “no ``` code block, all on one line”
⚠️ Where people go wrong
- A grand role (“You are a world-class expert…”) does not fix a missing output format.
- Asking the AI to explain its answer adds text and fails exact-match tests.
- Forgetting the {input} placeholder: the AI never sees the test data.
📌 Remember
- Input → rules → format → example → {input}
- Reply with ONLY …
- Repeats, ties, empty, capitals, rounding
- Try it on tricky inputs: Prompt lab