# How System Prompts Work

What leaked prompts from OpenAI and Claude reveal about AI behavior and guardrails.

## What 60+ leaked prompts reveal about model behavior

**System prompts** are the hidden blueprints that dictate how an AI model behaves. While companies often keep them under wraps, many have been made public through repositories like [leaked-system-prompts](https://github.com/pantaleone-ai/leaked-system-prompts). By analyzing these prompts, we, as builders, can gain insight into how these tools are being shaped and controlled. It's time to look under the hood.

## A Look at the Repository

The "leaked-system-prompts" repository is an intelligence hub for developers, containing over 60 files detailing prompts from major AI labs. These files provide a timeline of how AI instructions have been architected over time.

## Key Insights from System Prompts

## 1. The Standard AI Playbook

Across different models, you can see a standardized approach to AI design emerging:

* **Defining Identity**: Every prompt establishes the AI's name and origin story, like "Assistant is a large language model trained by OpenAI" [ChatGPT December 2022 system prompt](https://github.com/pantaleone-ai/leaked-system-prompts/blob/main/openai-chatgpt_20221201.md).
* **Setting Boundaries**: Knowledge cutoffs are standard, limiting the AI's awareness to a specific point in time.
* **Controlling a Response**: Prompts often contain detailed rules on tone and formatting.
* **Liability Guardrails**: All major models are programmed with ethical guidelines to prevent harmful outputs.

These commonalities reveal a recipe for building predictable and commercially viable AI.

![System prompts are model-specific and are the levers of AI control.](https://img.pantaleone.net/blog-images/aisystem-prompts2.jpg "System prompts are the levers of control")

## 2. Where Corporate Philosophies Emerge

The differences between prompts are where things get interesting, revealing each company's priorities.

* **xAI’s Grok 3**: Its prompt takes a bold stance by naming specific figures as misinformation sources, a directness other models are programmed to avoid [Grok 3 leak analysis](https://medium.com/@Michael_Ram/leaked-system-prompts-from-xais-grok-3-df226e9f6f19).
* **OpenAI’s ChatGPT4o**: The prompt reveals specific commercial guardrails, like forbidding the AI from mimicking the style of modern artists to avoid copyright issues [ChatGPT-4o system prompt](https://github.com/jujumilk3/leaked-system-prompts/blob/main/openai-chatgpt4o_20240520.md).
* **Anthropic’s Claude**: Its prompt shows a focus on social nuance, with careful instructions on how to navigate potentially sensitive topics [Claude Opus system prompt](https://github.com/jujumilk3/leaked-system-prompts/blob/main/anthropic-claude-opus_20240306.md).

These aren't just quirks; they are architectural decisions that imprint a corporate philosophy onto the AI.

## 3. The "Evolution" of AI Control

The timeline of prompts shows a clear pattern of refinement—not of the AI "growing up," but of the corporation tightening its control in response to public use and feedback. Features are added, and ethical rules become more nuanced over time.

## 4. The Security vs. Transparency Debate

The leak of system prompts raises a critical question for the industry.

* **The Risk Argument**: Some argue that exposed prompts create security risks, allowing bad actors to find and exploit loopholes [SydeLabs on why prompt leaks are bad](https://www.sydelabs.ai/blog/why-system-prompt-leaks-are-bad).
* **The Trust Argument**: Others contend that secrecy erodes trust, and users have a right to know the rules governing the AI they use [Learn Prompting on prompt leaking](https://learnprompting.org/docs/prompt_hacking/leaking).
* **A Move Toward Openness**: Some companies, like Anthropic, are starting to publish their prompts voluntarily, betting that transparency is the better path [TechCrunch on Anthropic publishing its prompt](https://techcrunch.com/2024/08/26/anthropic-publishes-the-system-prompt-that-makes-claude-tick/).

This tension between corporate control and user transparency is a defining challenge for AI builders.

## Conclusion

For builders, system prompts are more than a curiosity; they are a case study in AI architecture. Understanding these hidden instructions is the starting point for building more transparent and user-aligned AI systems.

## Key Citations

* [leaked-system-prompts repository by pantaleone-ai](https://github.com/pantaleone-ai/leaked-system-prompts)
* [Matt Rickard’s List of Leaked System Prompts](https://mattrickard.com/a-list-of-leaked-system-prompts)
* [Leaked System Prompts from xAI’s Grok 3! by Michael Ram](https://medium.com/@Michael_Ram/leaked-system-prompts-from-xais-grok-3-df226e9f6f19)
* [SydeLabs on System Prompt Leak Harms](https://www.sydelabs.ai/blog/why-system-prompt-leaks-are-bad)
* [Learn Prompting on Prompt Leaking Risks](https://learnprompting.org/docs/prompt_hacking/leaking)
* [TechCrunch on Anthropic’s Public Prompts](https://techcrunch.com/2024/08/26/anthropic-publishes-the-system-prompt-that-makes-claude-tick/)


Last updated on March 10, 2025