We often interact with AI models assuming they are just predicting the next word, but there is a massive layer of hidden instructions guiding every response. A recent discovery suggests that the complexity of these hidden "system prompts" is scaling up dramatically, potentially revealing the architecture for upcoming models like GPT 5.2. This incredible discovery comes from PerceptionGrand556, who managed to extract what appear to be the full system instructions for unreleased or internal test versions of OpenAI’s models.
When I read through these logs, I was genuinely surprised by the level of micromanagement involved in the model’s behavior. The author shared raw text that references a knowledge cutoff of August 2025 and a current date of January 2026, which hints that these might be future-proofed internal test builds. The prompts cover everything from a new tool called "Canmore" to highly specific protocols for handling email drafts and arithmetic. It is a fascinating look at how developers are using natural language to patch software limitations.
The Shift to Behavioral Scripting
The most significant takeaway from this find is how OpenAI is explicitly scripting the model’s "thought process" rather than relying solely on training data. The original poster revealed that the system prompt doesn’t just tell the AI who it is; it gives it a step-by-step flowchart for specific tasks. For instance, there is a massive section dedicated to "Writing Blocks," a UI feature specifically for emails. The model is instructed to plan the email in chat first, decide if a distinct UI block is needed, and then generate it. This proves that the future of AI interaction isn’t just about text generation; it’s about the AI understanding the interface it inhabits and using it to present information more cleanly.
📌 The "Writing Block" Protocol
The instructions for drafting emails found by the expert are incredibly specific and revealing about how we should be prompting current models. The system prompt explicitly tells the AI to consider providing multiple variants, such as "more formal," "more concise," or "more persuasive", when a user asks for help.
This is a massive insight into what the developers consider a "quality" response. They explicitly instruct the model to precede each email draft with a concise explanation of the variant’s intent. The goal is to separate the planning of the email from the execution of the draft. For us, this validates a specific prompting strategy: never just ask for an email. Ask for the plan first, then ask for three distinct variants with clear tonal shifts. The system prompt essentially forces the AI to act as a consultant before acting as a writer, ensuring that the final output isn’t just a generic block of text but a targeted communication tool.
💡 The "Digit-by-Digit" Safety Net
One of the most amusing yet practical parts of the post is the section on logic and riddles. The Reddit user highlighted instructions where the system is aggressively warned about trick questions. The prompt literally commands the AI to "pay close attention to the exact wording" and to "second-guess all assumptions," specifically for classic riddles.
Even more interesting is the instruction for math. We know LLMs are bad at arithmetic, and apparently, so does OpenAI. The system prompt forces the model to calculate "digit by digit" for any arithmetic query. It is forbidden from relying on memorized answers. This confirms that to get accurate math from an LLM, you cannot let it maintain a high-level overview; you must force it into a granular, step-by-step process. The prompt explicitly states that studies show the model makes mistakes without this breakdown. This is a clear signal that "Chain of Thought" isn’t just a fancy term; it is a mandatory safety rail for the newest models.
✅ The "Canmore" and Personality V2
The discovery also references a tool called "Canmore," which appears to be the internal codename for the Canvas feature (collaborative writing and coding). The instructions for this tool are distinct from the standard chat, emphasizing collaborative editing over simple generation.
Furthermore, the "Personality v2" section outlines exactly how the AI should sound: "Supportive thoroughness," "Lighthearted interactions," and "Confidence-building." It explicitly bans hedging phrases like "Would you like me to…" or "I can do that if you want." The developers are trying to kill the sycophantic, overly polite loops that earlier models fell into. They want the AI to be proactive: if the next step is obvious, the instructions say to just do it. This shift towards agency, acting without asking for permission for every tiny step, is a major clue regarding how the user experience is being tuned for future iterations.
How to Mimic the "GPT-5" Logic
Based on the instructions PerceptionGrand556 found, you can improve your current prompts by manually injecting these "system-level" constraints. You don’t need to wait for the update to use the logic.
1. Force the Math Protocol
Whenever you need calculations or logic, use the exact phrasing from the leak:
"Calculate this digit by digit. Do not rely on memorized answers. Work out the answer step-by-step before stating the final result."
2. The Email Consultant Workflow
Don’t ask:
"Write an email to my boss about missing the deadline."
Do ask:
"I need to email my boss about a missed deadline. First, determine the best content, tone, and structure. Then, provide three variants: one apologetic and formal, one solution-oriented and concise, and one that is casual. Precede each variant with an explanation of why you wrote it that way."
3. The Anti-Hallucination Riddle Solver
If you are asking complex logic questions, verify the output by adding:
"Pay close, skeptical attention to the exact wording of my query. Assume I might be using adversarial wording. Second-guess your first assumption before answering."
This leak gives us a rare manual on how the creators talk to their own creation. By adopting their language, we can get superior results right now!
If you want to read the full raw text of the system prompts to see the specific penalty clauses and tool definitions, you should check out the original discussion.
💡 FAQ & Troubleshooting
Can I use “Writing Blocks” for content types other than email?
No. While writing blocks effectively produce uncontaminated text artifacts in the UI, the current syntax rules strictly limit the variant attribute to "email". Attempts to use other variants are not supported by the current system prompts.
How do I enable the memory feature if the ‘bio’ tool is disabled?
The internal bio tool is explicitly disabled in the system instructions. To allow the model to remember personal details, you must navigate to Settings > Personalization > Memory and enable the feature there.
Why is the model calculating math problems digit by digit?
This is a mandatory behavior protocol. To prevent arithmetic errors, the system prompt explicitly instructs the model to calculate “digit by digit” and avoid relying on memorized answers or one-sentence responses for numerical questions.
What happened to the ‘browser’ tool?
The browser tool has been deprecated. The system now uses the web tool, which utilizes search() and open_url() commands to access up-to-date information.
I found current ChatGPT system prompts
byu/PerceptionGrand556 in