Good Examples and Bad Examples
Unit ID: PAW-M06-U01 Estimated active time: 20-25 minutes
Showing what to avoid
A bad example is often more informative than a good one, because it names a failure the tool would otherwise repeat.
> Here is a weak version and a strong version of the same summary. Match the strong one. The weak one fails because it states a cause the notes do not support.
The pairing is what teaches
A bad example alone is risky — it is still a pattern in the prompt, and patterns get imitated.
Always pair it with the corrected version and a one-line reason. The reason is what generalises; without it you have taught one specific fix.
A worked contrast
CONTRAST PAIR
WEAK
"Attendance dropped in week three, probably because of the weather.
Overall the course went well and learners were happy."
WHY IT IS WEAK
- "probably because of the weather" -- a cause not in the notes
- "learners were happy" -- no source; feedback covered 94 of 240
- "went well" -- unfalsifiable
STRONG
"Attendance fell from 186 to 141 in week three. The notes do not record a
reason. Feedback was collected from 94 of 240 learners; among those, 78 rated
the pacing positively. No conclusion is available for the other 146."
WHY IT IS STRONG
- every number is from the notes
- the gap is named rather than filled
- the limit of the feedback is statedDraw examples from real failures
The best bad examples are outputs you actually received and rejected.
Keep them. A prompt improved with a real failure it produced is much more likely to prevent that failure than one improved with an invented example.
Do not stack too many contrasts
Two contrast pairs are usually the limit before the prompt becomes mostly examples.
If you need more than two, the underlying problem is probably a missing constraint or an unsplit task rather than insufficient demonstration.
Mini practice
Find one AI output you rejected.
Write the contrast pair: the weak version, why it is weak, the strong version, why it is strong.
Add it to the prompt that produced it, and re-run.