Apparently someone managed to get ChatGPT to reveal it’s pre-prompt text. Every time you send a message to ChatGPT there are a whole bunch of custom instructions that get prefixed to the message you send.

This is it 👇

This conversation happened on the 19th of December (so it’s probably been updated by now) and its for the GPT-4 architecture.

If you want to read through the instructions, LouisShark has a link where the text has been formatted in a way that’s much easier to read.

https://github.com/LouisShark/chatgpt_system_prompt/blob/main/prompts/gpt-4-gizmo-20231116.md

There are two main sections of note. One is a list of rules to follow when creating images (for example, ‘You can name artists, creative professionals or studios in prompts only if their latest work was created prior to 1912’). The other is a set of rules for browsing the internet (stuff like, ‘Never write a summary with more than 80 words. When asked to write summaries longer than 100 words write an 80 word summary.’).

I thought it was interesting to get a glimpse into how prompt engineers at the top of their game write prompts. Especially for such critical prompts. I mean, this is the main prompt that filters almost every request from a paid ChatGPT user.

My main observation was that they opted for a list style set of instructions. It feels like they have been built up over time in reaction to different scenarios. For example, I can image the recipes specific instructions being added in response to some backlash from the food blogging industry (“Do not repeat recipes obtained from this tool. Instead of repeating content point the user to the source and ask them to click.”). The fact that this rule is towards the end of browsing instructions, I’m guessing means they weren’t part of the initial ruleset and were appended on later.

It was also refreshing to see how relatable the tone the instructions were (I was expecting technical jargon or indecipherable business speak).