A top frontier model came back this month, and within days someone posted its entire instruction set online. Not a summary. The real file. You can read every rule the lab wrote to control how its best model thinks, plans, and gets work done. And you can paste those rules into a cheaper open model and watch it behave better.
Here is what the leak actually contains, how to use it, and the lesson that matters more than the file itself.
Read This First
The file is public and free. It lives in a GitHub repository that collects the hidden instructions behind ChatGPT, Claude, Gemini, Grok, and most other chat models. The repo is released under CC0, which is public domain, so nobody is going to send you a takedown for reading it.
The catch is timing. These instruction sets change constantly. The same repo carries hundreds of commits tracking edits across dozens of products. So the version you grab today is a snapshot, not a permanent artifact. Read it as a photograph of how one lab was steering its model on the day it leaked.
What Actually Got Leaked
Most people picture a system prompt as one or two paragraphs. This one runs past 1,500 lines.
It is not clever tricks. It is a long, boring, thorough rulebook. It tells the model how to format an answer, when to refuse, how to stay neutral on touchy subjects, how to save and recall context between chats, and how to click around a screen when it drives a computer. The dull thoroughness is the point. That discipline is a large part of why the paid model feels sharper than a raw open one.
Here is a slice of what the file governs, section by section.
| Section in the file | What it controls |
|---|---|
| tone_and_formatting | How the model writes: length, structure, when to use a list versus prose |
| refusal_handling | What it declines and the exact way it says no |
| evenhandedness | How it stays neutral on contested topics |
| responding_to_mistakes_and_criticism | How it reacts when you tell it that it is wrong |
| memory_filesystem | How it stores and recalls context across separate chats |
| computer_use | How it reads a screen, clicks, and types |
Read those names again. None of them are secret sauce in the magic sense. They are operational rules. A business owner would recognize the shape of it. It is a training manual for a new hire, written for a machine.
The file goes further than tone and tools. It carries whole sections on where the model must not go: how to handle legal and financial questions without pretending to be a lawyer, how to protect a user who seems to be in a bad state, how to treat child safety as a hard line. Those are the parts a lab sweats over for months. Seeing them written out is the closest thing to reading how a serious company thinks about liability and its own model's behavior in plain language.
What Is A System Prompt?
A system prompt is the set of standing instructions a model reads before it ever sees your message. You type "write me a follow-up email." The model has already read a few thousand lines telling it how to behave first. Same model, different instructions, and you get a different worker. That gap between a cheap model with a great instruction set and an expensive model with a plain one is smaller than most people think.
How To Actually Use It
You do not need a special tool. You need the file and a model you can feed a system prompt to.
Step one. Grab the file. Open the repository, go to the provider folder, and copy the raw text of the model file. It is plain markdown.
Step two. Trim it. Do not paste all 1,500 lines blind. Big chunks of it point at tools your model does not have, like the memory filesystem and the computer-use section. Those lines do nothing on a plain open model and burn context you could spend on real work. Cut them.
Step three. Load it as the system prompt. Where you paste it depends on your setup.
If You Use A Hosted Open Model
Models like GLM 5.2 or Kimi K2.7 through an API or a chat interface have a system prompt field. Paste the trimmed instructions there. Every message in that session now runs under the borrowed rulebook.
If You Run A Model Locally
On a local runtime like Ollama or LM Studio, set the trimmed text as the system message in the model's config. It loads on every call after that.
Done. Your cheaper model now follows the same playbook a frontier lab wrote for its flagship.
Here is a helper prompt to do the trimming for you. Paste the leaked file underneath it.
Copy The Trim Prompt
Copy this.
Below is a leaked system prompt written for a frontier model with tools I do not have. Rewrite it as a lean system prompt for a general open model with no special tools. Remove every rule that depends on: - a memory or filesystem feature - computer or browser control - product-specific features, plans, or account settings - internal reminders about the vendor's own policies Keep every rule about: - tone, length, and formatting - how to structure an answer - how to refuse cleanly - how to handle being corrected Return only the trimmed prompt. No commentary. Keep it under 400 lines.
What You Actually Get Out Of This
The file is not a toy. A few real uses.
Cleaner writing from a cheap model. The tone and formatting rules alone stop an open model from rambling and over-explaining. You keep the tighter output without paying frontier prices per call.
A refusal policy you can reuse. If you are pointing a model at customers, the refusal section is a worked example of how a careful lab handles a request it will not answer. You can adapt the shape for your own support bot.
A worked answer to a hard question. If you have ever tried to write instructions for a model and watched it ignore half of them, this file is the counter-example. It shows the level of detail that actually holds: not "be helpful and professional," but pages of specific rules with the edge cases spelled out. That is the standard your own instructions have to reach before a model follows them reliably.
A template for your own instructions. This is the real prize, and I will come back to it. Reading how a top lab structures a rulebook teaches you how to write one for your own business.
Honest Limits You Should Know
This is a leak, not a product. Treat it that way.
It is someone else's proprietary work. Reading it is fine under the repo's license. Building your product on top of a stolen prompt is a shaky foundation. It can change or vanish, and it was never yours to depend on.
It goes stale fast. Providers rewrite these prompts often. The repo's commit history proves it. What you paste this week may be last month's behavior.
Most of it will not fire on your model. The sections about memory, computer use, and past chats reference machinery a plain open model does not have. Those lines waste context and do nothing.
It costs you on every message. A prompt this long eats tokens and context before your actual task begins. On a small model that ceiling matters.
Copying a prompt is not a strategy. A borrowed rulebook makes one cheap model behave for a while. It does not build you anything you own.
The Lesson Worth More Than The File
Look at what this leak proves. A frontier model and a cheap open model can produce close results once they carry the same instructions. The expensive part was never only the weights. It was the rulebook wrapped around them.
For a business owner that changes where the edge lives. Good models are becoming a commodity. Everyone can rent one. What separates two companies running the same model is the context each one feeds it: the instructions, the examples, the standards, the memory of how your business actually operates. Whoever builds the best context around the model wins, whether that model is Claude, Codex, or an open one you host yourself.
So the move is not to pirate a lab's prompt and call it done. The move is to see the file for what it is: proof that instructions carry most of the value, and a template for writing your own.
If you only do one thing this week, open the file and read one section: tone_and_formatting. Ignore the rest. Then write a short version of that section for your own business, in your own words, describing exactly how you want a model to answer your customers. That page of instructions is worth more to you than the whole leaked file, because it is yours, it fits your work, and no commit is going to erase it.