Leaked system prompts of popular LLMs and what they tell us

In the ever-changing world of large language models or LLMs, the system prompts guide AI behavior under the hood, one of the many reasons why different models have different reasonings for the same question.

A repository on GitHub, jujumilk3/leaked-system-prompts (opens in a new tab), lists reported system prompts of some of the most popular LLM products, offering a glimpse into how platforms like OpenAI’s ChatGPT, Anthropic’s Claude, GitHub’s Copilot, and more may have been instructed to operate.

This blog consists of the most interesting system prompts, cherry-picked to give you actionable advice on crafting the perfect prompt for each model.

Why system prompts matter#

As quoted above, system prompts shape an LLM’s tonality, reasoning, and boundaries. Each AI Model is marketed differently to provide generalized advice or perform a specific task with perfection. By deconstructing and understanding the system message of an LLM, we can curate better prompts that are more definitive and respect an LLM’s boundaries to maximize the model’s capabilities.

Sources and dates: what these examples establish#

The dates below come from the archive filenames, not confirmed vendor release dates. These are community reports; a model repeating instructions doesn't establish their authenticity or completeness. Read the examples as historical snapshots and the advice as my interpretation of them.

For a vendor-published source, Anthropic documents Claude's core system prompts by model and date (opens in a new tab), including Claude Sonnet 3.5 (opens in a new tab). Anthropic says these prompts apply to its web and mobile apps, and these updates do not apply to the API. Product, model, and date all matter when comparing prompts.

Key Findings#

Let’s talk about five popular LLM products through their 2024 snapshots. Click on the headings to read the archived system prompt for each one!

1. GitHub Copilot (opens in a new tab)#

Archive filename date: Sep 30, 2024

GitHub Copilot is an AI-powered code completion tool that helps developers write code faster and more efficiently.

Key Takeaways from GitHub Copilot#

• GitHub Copilot prioritizes code completion and developer assistance.
• It emphasizes concise and relevant responses and avoids verbosity unless requested.
• The snapshot describes tools for retrieving repository files and security alerts.

Writing the best prompts for GitHub Copilot Chat#

• Be specific and clearly describe the task to be performed. Though Copilot isn’t designed to be verbose, it expects verbosity from the user and performs best if the language, and framework. and desired output are explicitly mentioned.
• Should you require additional context, ask for explanations by simply appending “Explain why.” for an in-chat explanation, or “Add inline comments” which is self-explanatory.

What’s interesting about GitHub Copilot?#

The system prompt incorporates guardrails and security-alert tools intended to support developer assistance. These instructions don't guarantee secure code, so generated changes still need review and testing.

2. xAI Grok (opens in a new tab)#

Archive filename date: Oct 3, 2024

Grok is a generative AI model that uses natural language processing to answer questions, perform creative writing, and more.

Key Takeaways from Grok#

• This Grok 2 snapshot emphasizes helpfulness, truthfulness, and questioning popular narratives.
• The system message explicitly encourages engaging with controversial questions instead of refusing them outright.

How to write the best prompts for Grok#

• Grok’s strength lies in its nuanced reasoning, so it’s recommended to give it a go for asking open-ended questions that leverage its philosophical and ethical reasoning.
• Encourage the model to present different viewpoints of a situation for a multi-perspective analysis that its capable of performing.

What’s interesting about Grok?#

Grok’s system prompt emphasizes on questioning popular narratives, suggestive that it’s designed for open-ended discussions with the user than providing a one-time resolution.

3. Anthropic Claude 3.5 Sonnet (opens in a new tab)#

Archive filename date: Nov 22, 2024

Anthropic's Claude is a family of large language models (LLMs) that power AI chatbots. Claude is designed to be safe, accurate, and secure.

Key Takeaways from Claude#

• Claude’s system prompt prioritizes alignment with user intent while maintaining a friendly, conversational tone.
• It’s instructed to be sensitive and empathetic, as highlighted by the following line: Claude is always sensitive to human suffering, and expresses sympathy, concern, and well wishes for anyone it finds out is ill, unwell, suffering, or has passed away.

How to write the best prompts for Claude#

• Be conversational and frame your questions naturally to take advantage of the model’s ability to simulate tonality.
• Establish context in a highly detailed manner for creative tasks. Claude appreciates verbosity.

What’s interesting about Claude?#

The system message includes instructions for engaging conversation, making it an interesting example for creative writing and empathetic communication.

4. Cursor IDE (Built on Claude Sonnet) (opens in a new tab)#

Archive filename date: Dec 24, 2024

Cursor AI is an innovative AI-powered code editor designed to streamline and enhance the coding process.

Key Takeaways from Cursor#

• Similar to Copilot, Cursor is optimized for IDEs to focus on coding assistance and debugging.
• The system message heavily emphasizes on context awareness for file structures and existing code.

How to write the best prompts for Cursor#

• Reference file context while asking questions related to the codebase by explicitly mentioning the respective files or file structure as a whole.
• While debugging issues, request step-by-step explanation for verbosity and in-depth analysis.

What’s interesting about Cursor?#

Cursor’s system prompt highlights the ability to interact with multiple files and project contexts at once, which makes it a powerful tool for AI-assisted coding.

5. OpenAI ChatGPT GPT-4o (opens in a new tab)#

Archive filename date: May 20, 2024

OpenAI ChatGPT is a chatbot that uses artificial intelligence (AI) to generate human-like responses.

Key Takeaways from ChatGPT#

• ChatGPT heavily emphasizes on versatility, allowing for both structured and creative outputs.
• The system message is optimized to balance formal knowledge delivery with conversational adaptability.

How to write the best prompts for ChatGPT#

• ChatGPT combines the best of both worlds; technical knowledge and creative reasoning. Experiment with the model’s versatility in your prompt.
• Curate multi-message prompts that build over context to maximize the model’s conversational capabilities.

What’s interesting about ChatGPT?#

ChatGPT’s archived prompt includes instructions for browser, Python, and image-generation tools, offering a glimpse into how different capabilities were brought together.

Tool Calling in Generative AI#

Tool calling emerges as a pivotal feature in these LLM system prompts, enabling enhanced interactivity and functionality across platforms. Several of the snapshots include tool descriptions; the surrounding application still has to implement and execute those tools.

• Claude 3.5 Sonnet has separate text-only and image-enabled instructions in this archive; this snapshot doesn't document external-tool access.
• ChatGPT GPT-4o explicitly references tools like dalle for image generation, showcasing its intended versatility in creative and technical workflows.
• Meanwhile, GitHub Copilot Chat focuses tool use on developers, with repository search, file retrieval, and security-alert tools.

Read more about tool or function calling in AI on OpenAI’s official documentation (opens in a new tab).

Ending Notes#

The archived system prompts of the LLMs mentioned above provide clues into how these products were instructed to operate. By understanding how things work under the hood, users can tailor their prompts for better, more optimized results.

With these recommendations, experiment, refine your prompts, and unlock the full potential of Generative AI!