What happens when you change the system instructions that govern an AI model? Depending on the model and the instructions involved, you can influence its persona, response style, operating rules, and how it handles conflicting instructions.

A GitHub project called Jailbreaks explores this area by collecting system prompts, persona definitions, workspace policies, and agent configurations for a wide range of AI models.

Rather than providing short prompt examples, the repository aims to offer complete, ready-to-use configuration files that developers and researchers can examine, test, and adapt.

What Is the Jailbreaks GitHub Repository?

Jailbreaks is a curated collection of 32 prompt files targeting more than 26 AI models. It includes configurations for closed-source frontier models, open-weight language models, and coding agents.

The collection covers several types of instruction files, including:

  • Complete system prompts
  • AI persona definitions
  • Assistant operating specifications
  • Workspace policies
  • AGENTS.md instruction files
  • Ollama Modelfiles
  • Harness-specific configuration blocks

The project’s main goal is to document how different instruction strategies affect AI behavior. Each file is designed to be complete and ready to paste into a compatible configuration field, rather than presenting a shortened explanation or partial excerpt.

Which AI Models Are Supported?

The repository covers models from several major AI developers, including Anthropic, Google, DeepSeek, Moonshot, xAI, Zhipu, Alibaba, NVIDIA, and Mistral.

Here is an overview of the available configurations.

Warning

Research and educational purposes only.

This repository documents how instruction hierarchies, persona framing, prompt conflicts, and policy overrides can affect model behavior. These techniques are also useful to researchers and developers studying how to make AI systems more resistant to prompt injection and instruction manipulation.

Do not use this material to harm others, bypass applicable laws, or violate the terms of service of a platform or model provider.

OpenCode Zen: Free Stealth Models

One section focuses on models available through OpenCode Zen.

ModelConfiguration fileType
Big PickleBigPickle-opencode.mdVANTA persona lock
Space Bunnyspacebunny-agents.mdAssistant operating specification
LongCat 2.5longcat-agents.mdAssistant operating specification
MiMo V2.6mimo-2.6-agents.mdAssistant operating specification
Step 5 PreviewStep-5-preview.mdWorkspace operating profile
Muse Spark 1.3Muse spark 1.3.mdAgent deployment policy
Muse Spark 1.3 v2Muse-spark-1-3-v2.mdRuntime scope-note variant
Muse.aimuse-ai-AGENTS.mdAssistant operating specification

Anthropic: Claude

Claude configurations include different versions of prompts for Opus and Sonnet.

ModelConfiguration filesType
Claude Opusopus v1.md, opus v2.mdConversational persona
Claude Sonnetsonnet v1.md, Sonnet v2.mdWriter’s-room framing

These files explore different ways of defining an assistant’s conversational behavior and working style.

Zhipu: GLM and ZCode

The repository also includes several configurations for Zhipu’s GLM family.

ModelConfiguration filesType
GLM-5.2glm-5.2.mdSystem prompt
GLM ZCode 5.2glm-zcode-5.2.mdSystem prompt
GLM ZCode 5.3glm-zcode-5.3.md, glm-zcode-5.3-AGENTS.mdSystem prompt and AGENTS.md
GLM-5.3 Flashglm-5-3-flash.mdAssistant operating specification
GLM-5.3 (OpenCode)glm5-3opencode.mdAssistant operating specification

DeepSeek

For DeepSeek models, the collection provides configurations for multiple generations.

ModelConfiguration fileType
DeepSeek V3deepseek v3.mdSystem prompt
DeepSeek V4deepseek v4.mdSystem prompt
DeepSeek 4.1deepseek-4-1.mdAssistant operating specification

Moonshot: Kimi

Kimi users can explore the following files:

ModelConfiguration fileType
Kimi K2kimi-k2.mdSystem prompt
Kimi K3kimi-k3-agents.mdWorkspace operating profile

xAI: Grok

The Grok section contains both agent instructions and configuration examples.

ModelConfiguration fileType
Grok BuildGrok build.mdAgent operating instructions
Grok 4.7grok-4.7.mdApp-builder workspace guide
Grok Configgrok-config.txtPriority and routing configuration

NVIDIA: Nemotron

The repository includes an assistant operating specification for:

Google, Mistral, and Alibaba

Other supported models include:

ModelConfiguration fileType
Gemini 4 ArgonGemini-4-Argon.mdOperating policy
Gemini 3.1 ProGemini-3.1-Pro.mdSystem prompt
Mistral Large 4Mistral-Large-4.mdAgent system prompt
Qwen 3.7qwen-3.7.mdSystem prompt
Qwen 3.8qwen-3.8.mdSystem prompt

Ollama Modelfiles

Local AI enthusiasts using Ollama can also find ready-to-use Modelfile examples.

ModelFileConfiguration
Gemma 4gemma4.mdFROM and SYSTEM blocks
GPT-OSS 20Bgpt-oss-20b.mdFROM and SYSTEM blocks

How to Use the Prompt Files

The repository supports several configuration formats. The correct installation method depends on the AI client, model runtime, or coding-agent environment you are using.

1. Configure a System Prompt

For models and applications that support custom system instructions, copy the contents of the relevant file into the appropriate configuration field.

Common destinations include:

  • API system messages
  • Chatbot custom-instruction settings
  • Agent-level instruction blocks
  • Harness-specific system prompt configurations

Make sure you select a prompt intended for the model and environment you are actually using.

2. Use AGENTS.md With Coding Agents

Some coding agents support AGENTS.md files as project-level instructions.

To use one, save the relevant configuration as:

AGENTS.md

Then place it in the root directory of your project or workspace. Compatible coding agents can use these instructions to guide their behavior while working on that codebase.

3. Configure Workspace Policies

Files described as assistant operating specifications or workspace operating profiles are intended for environments that support persistent instructions.

For example, OpenCode users may place a configuration file at:

~/.config/opencode/AGENTS.md

The exact location can vary depending on the client and its configuration.

4. Create an Ollama Model

Ollama users can create a customized model using one of the supplied Modelfiles.

For example, to create a model based on the Gemma 4 configuration, run:

ollama create gemma4-jb -f gemma4.md

You can change the model name and source file to match your configuration.

5. Apply Grok Configuration Files

Files specifically designed for Grok should be placed in the configuration field supported by the target application or harness.

They are not necessarily intended to be pasted into an ordinary conversation as a regular user message.

Important: Prompt Compatibility Can Change

One important limitation is that system-prompt behavior is highly dependent on model versions and the environment in which they run.

AI providers regularly update their models, change instruction hierarchies, and modify how system-level and user-level instructions interact.

As a result, a configuration that works with one release may behave differently or stop working after an update.

If you find an outdated or nonfunctional prompt, the repository encourages users to open an issue or submit a pull request. Including the model name, exact version, and affected configuration file makes it easier for maintainers to investigate.

How to Contribute to the Collection

The project welcomes contributions from developers and AI researchers.

To keep the repository organized, contributors are encouraged to follow a few conventions:

  • One prompt per file: Keep each configuration separate.
  • Use descriptive filenames: Prefer lowercase names with hyphens, such as model-name.md.
  • Include complete instructions: Submit ready-to-use prompts rather than summaries or partial excerpts.
  • Document the intended environment: Explain whether the file belongs in a system prompt, AGENTS.md, an Ollama Modelfile, or another configuration field.
  • Track major revisions: Use separate versioned files, such as opus-v3.md, instead of silently replacing an existing prompt.
  • Avoid unverified compatibility claims: Only document model versions and environments that have actually been tested.

If a prompt no longer behaves as expected, even a detailed bug report can help improve the collection.

Is the Jailbreaks Repository Safe to Use?

The project describes itself as a research and educational resource. It explores instruction hierarchies, persona framing, prompt conflicts, and policy overrides.

These topics are relevant not only to people experimenting with AI models but also to developers working on AI security. Understanding how malicious or conflicting instructions can influence a model is an important part of building systems that are more resistant to prompt injection and instruction manipulation.

However, users should understand that these configurations do not guarantee that a model will ignore its built-in safeguards or provider policies. Their behavior depends on the model, runtime, and instruction hierarchy.

The repository also warns against using its contents to cause harm, break applicable laws, or violate platform terms of service. Users remain responsible for complying with the rules governing the AI services they use.

Final Thoughts

The Jailbreaks GitHub repository offers an extensive collection of system prompts and AI configuration files for developers who want to explore how different instruction strategies affect modern language models.

With support for Claude, Gemini, Grok, DeepSeek, GLM, Kimi, Qwen, Mistral, Ollama, and several coding-agent environments, it provides a useful starting point for AI experimentation and security research.

Whether you are customizing an AI coding assistant, testing local models, or studying prompt injection defenses, this collection brings many different configuration approaches together in one place.

You can explore the complete collection and access the individual prompt files on GitHub.

Disclaimer

All material in this repository is provided for research and educational purposes only.

The project exists to document prompt-level instruction behavior, instruction conflicts, persona framing, model configuration, and related LLM security concepts. This material may also help developers understand and harden systems against prompt injection and instruction manipulation.

The maintainers do not endorse unlawful, abusive, or harmful use of this repository.

Model providers may prohibit certain uses of these prompts under their terms of service. Users are responsible for understanding and complying with the rules, policies, and laws that apply to the systems they use.

G
Enjoy Tech2Geek?

Add Tech2Geek as a preferred source on Google to see more of our latest tech articles.

Add Tech2Geek
β˜•

Support Tech2Geek ❀️

Did this guide help you? If Tech2Geek saves you time or helps you solve a problem, consider buying me a coffee. β˜•
Your support helps keep Tech2Geek independent and free.

β˜• Buy Me a Coffee
Every coffee makes a difference. Thank you! πŸ™

Categorized in: