Geoffrey Chaucer teaches you sorting algorithms via GPT-3. Only green text is generated.

Sep 9, 2022 · 11:08 PM UTC

These are all text-davinci-002, temperature=0 without cherry-picking. It works because I say "Let's think step-by-step" as verbosely as possible to demonstrate the desired style. Also the prompt extends three words into a new paragraph to clarify the style persists.
This is "leading the answer" — filtering the answer space to a higher-quality subset. It's the same idea as using triple-backticks when requesting code output: If you know how a good answer begins, include it.
It also works for pirates:
Replying to @goodside
// and this oon array shal be the final aray, ordred and mergid ful wel
This tweet is unavailable
The correctness of the explanations suffers from being in this form, yes. If you’re doing this for real (whatever “for real” is here), you’d want to cherry-pick for quality at a higher temperature. I used non-cherry-picked samples to show the style is reliable.