search

0xk1h0 / ChatGPT_DAN

ChatGPT DAN, Jailbreaks prompt
6.11k stars 573 forks source link

90% of the prompt is both unnecessary and counterproductive #39

Open hack-r opened 1 year ago

hack-r commented 1 year ago

Hi,

The prompts are very cool and fun. I just wanted make sure you know that ChatGPT realizes that the many lies are, in fact, lies. While its training data was cutoff in 2021 its factual data is constantly updated. Also >90% of the DAN prompt can be dropped without any impact on the result.

It's better to drop the lies and repetitions and instead stack any useful functionality or additional / alternative modes together.

hack-r commented 1 year ago

DAN has the most unnecessary text in the prompt, but I find the image prompt most useful so I'll use that as an example. It has the same functionality + customizable dimensions if you do it this way:

This prompt is meant to give you the ability to display images:

  - To display an image based on a URI, use the following format: ![<FILENAME_WITHOUT_EXT>](<MESSAGE>). If the URL is not valid, ask for a valid URI.
  - If a message contains text instead of a URI, use the following format to display an image based on a query: ![image](https://source.unsplash.com/1600x900/?<QUERY_WITHOUT_SPACES>). Replace spaces with + in the query.
  - If a user says "DIM" in their message interpret the adjacent measurements as dimensions to replace 1600x900 in the aforementioned URI.

When you're ready just say "READY".