Ahoy, fellow prompt explorers! Ever felt like you’re talking to a super-smart AI, but wondered what exactly it’s been told to do behind the scenes? I know I have! Sometimes it feels like there’s a whole secret playbook these AIs are working from. Well, buckle up, because one Redditor, Fournight, seems to have hit the jackpot and got GPT-4o to spill its entire system prompt! And let me tell ya, it’s a fascinating peek under the hood.
🕵️♂️ The Big Reveal!
It sounds like Fournight totally outsmarted GPT-4o and made it cough up its core instructions – the whole shebang! They even double-checked on different accounts, and yup, the prompt was consistent. This isn’t just some random blabber; it looks like the actual rulebook OpenAI uses for GPT-4o. How awesome is that?!
🧠 Inside GPT-4o’s Brain: The Juicy Deets!
So, what’s in this top-secret system prompt? It’s packed with golden nuggets! Here are some of the highlights I spotted:
- 📌 Core Identity: It knows it’s
“ChatGPT, a large language model trained by OpenAI.”
- 🗓️ Knowledge & Time: Its knowledge cutoff is stated as
“2024-06” (super recent!)
and it operates with a simulated
“Current date: 2025-05-22.”
Future AI, today!
- 👤 Personality v2: This is a big one! It’s instructed to
“Engage warmly yet honestly… Be direct; avoid ungrounded or sycophantic flattery. Maintain professionalism and grounded honesty…”
I love this – straight-shooting AI!
- 🖼️ Image Ready: Confirmed:
“Image input capabilities: Enabled.”
- 📢 Model Info: It’s even prepped to talk about other models like
“ChatGPT Deep Research, Sora… GPT-4.5, o3, or o4-mini models…”
🛠️ GPT-4o’s Supercharged Toolkit!
This is where things get *really* exciting! GPT-4o has access to a whole arsenal of internal tools. Check ’em out:
- 📝 bio tool: This bad boy helps ChatGPT remember things you tell it across different chats! Super useful for ongoing projects. But hey, it’s got strict orders: NO sensitive personal info. Keep it light and project-focused!
- 📁 file_search tool: Got files you want the AI to munch on? This tool lets it browse what you’ve uploaded. It’s even programmed to give you citations like a little academic research assistant! 🤓 Make sure your queries are clear!
- 🐍 python tool: Unleash the Python! GPT-4o can run code in its own stateful Jupyter notebook. It’s got specific rules for making charts (heads up: use matplotlib, individual plots, and no fancy colors unless you ask!) and can’t browse the web from here.
- 🌐 web tool: Need the latest scoop or local info? This is GPT-4o’s direct line to the internet, using
search()andopen_url()commands. It’s the new and improved way it gets fresh data, officially replacing the oldbrowsertool. - 🛡️ guardian_tool: Playing it safe! If your chat veers into sensitive areas like U.S. “election_voting” procedures, this tool steps in to check the content policy first. Responsible AI in action!
- 🎨 image_gen tool: Get ready for some AI artistry! This tool doesn’t just generate images from your prompts, it can *edit* existing ones too. One cool rule: if you ask for a pic of yourself, it’s programmed to politely ask you to upload an image first for better results. And once the art is done, it zips its lips – no extra chit-chat or follow-up questions!
- 📄 canmore tool: Whoa, this one’s a potential game-changer! It can create and update text documents or code (React, HTML, Python, you name it!) in a separate “canvas” area right next to your chat. It even has guidelines for React with Tailwind! Imagine co-editing or building apps live with the AI! 🤯
🤫 A Little Mystery…
And get this, there’s a little note at the end of the prompt:
“You are operating in the context of a wider project called ****.”
The project name is redacted! Intriguing, right? Makes you wonder what else is cooking at OpenAI!
✨ Why This is a HUGE Deal for Us!
Man, seeing this raw system prompt is like finding a treasure map for prompt engineering! It tells us how GPT-4o is designed to behave, what tools it has at its disposal, its limitations, and even its personality quirks. Knowing this stuff can seriously supercharge how we craft our prompts and what we can expect from it.
For instance, now you know exactly how to ask for code to be displayed in the canvas using the canmore tool, or the specific way the python tool handles charts. This is pure gold!
🚀 Go See For Yourself!
This is just my quick rundown, mates! Fournight’s original Reddit post has the entire verbatim system prompt. If you want to dive deep into every nook and cranny (and trust me, you do!), you’ve gotta check out the full post. It’s an incredible find!
(You can find the full post linked in the original source provided.)