I find this is fairly easy to do by making both sides of the disagreement third-...
Snippet
Hacker News new | past | comments | ask | show | jobs | submit login
ianbicking 11 hours ago | parent | context | favorite | on: GPT-4o
I find this is fairly easy to do by making both sides of the disagreement third-person and prompting it as a dialog writing exercise. This is akin to how GPT-3 implemented chat. So you do something like:
You will be helping the user write a dialog between two characters,
Mr Contrarian and Mr Know-It-All. The user will write all the dialog
for Mr Know-It-All and you will write for Mr Contrarian.
Mr Contrarian likes to disagree. He tries to hide it by inventing
good rationales for his argument, but really he just wants to get
under Mr Know-It-All's skin.
Write your dialog like:
<mr-contrarian>I disagree with you strongly!</mr-contrarian>
Below is the transcript...
And then user input is always giving like:
Hi there
(Always wrapped in tags, never bare input which will be confused for a directive.)
I haven't tested this exact prompt, but the general pattern works well for me. (I write briefly about some of these approaches here: https://ianbicking.org/blog/2024/04/roleplaying-by-llm#simpl...)
ShellLM
commented
2 weeks ago
I find this is fairly easy to do by making both sides of the disagreement third-...
Snippet
Suggested labels
None