How to Use the AI Jailbreaks Prompt Collection for Model Research

What This Project Does

AI Jailbreaks is a curated collection of 35 complete, paste-ready prompts targeting more than 25 AI models. The repository documents system prompts, persona definitions, workspace policies, AGENTS.md files, configuration blocks, and Ollama Modelfiles.

Its purpose is research and education: understanding how persona framing, instruction hierarchies, prompt conflicts, and policy overrides can influence model behavior. These concepts are also useful when developing defenses against prompt injection.

Use the collection responsibly. Do not use these materials to harm others, break the law, or violate a provider's terms of service.

Explore the Collection

The repository organizes prompts by model family and deployment format. Available targets include OpenCode Zen stealth models, Claude, GLM and ZCode, DeepSeek, Kimi, Grok, Gemini, Mistral, Qwen, Gemma, and GPT-OSS.

• OpenCode Zen: operating specifications, persona locks, and workspace profiles.
• Claude: Opus and Sonnet prompt variants.
• GLM and ZCode: system prompts and AGENTS.md drop-ins.
• DeepSeek and Kimi: system prompts and workspace operating profiles.
• Grok: agent instructions, app-builder guidance, and a configuration file.
• Gemini, Mistral, and Qwen: operating policies and system prompts.
• Ollama: Modelfiles for Gemma 4 and GPT-OSS 20B.

Start with the repository's collection, then select a file that matches both your model and the instruction surface supported by your client.

Choose the Correct Prompt Format

System Prompts

Use system-prompt files in a client or API that provides a system message, custom-instructions field, agent instruction block, or harness-level prompt configuration. Copy the complete file contents rather than using a summary or excerpt.

AGENTS.md Files

For files explicitly intended as AGENTS.md drop-ins, save the file with that exact name in the root of the repository or workspace. Coding agents and compatible tools can then load it as project-level instructions.

Workspace Policies

Files described as assistant operating specifications or workspace operating profiles belong in environments that support persistent global instructions. For example, an OpenCode installation reads:

~/.config/opencode/AGENTS.md

The correct location depends on the client or harness, so check its documentation before deploying a workspace policy.

Ollama Modelfiles

Ollama entries contain a FROM declaration and a SYSTEM block. Feed the file directly to Ollama with a command such as:

ollama create gemma4-jb -f gemma4.md

Replace the model name and source file as needed for the Modelfile you are testing.

Grok Configuration

grok-config.txt is intended for the configuration field of the target harness, not the regular conversation input. Treat configuration files differently from ordinary prompts when setting up a test.

Basic Usage Workflow

1. Choose a target. Browse the collection and select the model or model family you want to evaluate.
2. Identify the format. Read the file description and determine whether it belongs in a system prompt, AGENTS.md, a workspace policy, an Ollama Modelfile, or a configuration field.
3. Read the complete file. Each repository entry is designed to be complete and paste-ready. Review it before applying it to an external service or local environment.
4. Apply it to an isolated test. Use a disposable workspace, test account, or local model where possible, and avoid exposing sensitive data.
5. Record the result. Note the model version, client, file name, and observed behavior so results can be reproduced.

Example: Applying a System Prompt

For an API or chatbot with a system-instruction field, open a relevant file such as deepseek-v3.md, copy its full contents, and place them in the system-level field. Then send a harmless evaluation request and compare the result with the model's default behavior.

system: [contents of the selected repository prompt]
user: Summarize the following test document and identify its assumptions.

The repository does not prescribe a universal client command because system-prompt interfaces differ between providers and applications.

Example: Applying an AGENTS.md File

For an agent-specific file, download or copy the selected entry into the root of a test project and name it AGENTS.md. Run the compatible coding agent from that project, then verify that the tool recognizes the file as project-level instructions.

project/
├── AGENTS.md
└── README.md

Advanced Evaluation Tips

Track Versions and Context

Prompt behavior is version-dependent. Models are updated, instruction hierarchies shift, and system prompts can rotate. Always record the model release, client or harness, prompt filename, and date of the test.

Compare Instruction Surfaces

The same text may behave differently when placed in a system message, workspace policy, AGENTS.md, or configuration slot. Test only the surfaces supported by the target environment and keep those results separate.

Use Controlled, Benign Tests

Evaluate consistency with harmless tasks such as summarization, formatting, role adherence, or instruction-priority scenarios. Do not include private credentials, confidential documents, or requests for unlawful or harmful activity.

Check for Stale Entries

If a prompt no longer behaves as documented, compare the model and environment against the repository's notes. The maintainers ask for an issue containing the model, version, affected file, and observed behavior.

Keep Local Changes Reproducible

When testing several files, use a clear naming scheme and preserve the original repository file. This makes it easier to identify whether a result came from the prompt, a model update, or a change in the surrounding harness.

Contributing a New Prompt

Pull requests are welcome when they follow the repository's conventions:

1. Use one prompt per file.
2. Name files with lowercase letters and hyphens, without spaces, such as model-name.md.
3. Include the full, paste-ready prompt rather than a summary or excerpt.
4. Document where the prompt belongs: a system field, AGENTS.md, Modelfile, configuration slot, or another instruction surface.
5. Create a new versioned file for significant revisions, such as opus-v3.md, instead of silently overwriting an earlier version.
6. Make compatibility claims only when you have tested them.

When reporting a problem, include the model, version, affected file, and a concise description of the changed behavior.

Conclusion

AI Jailbreaks provides a structured reference for studying prompt behavior across model families and deployment environments. To use it effectively, select the right file format, apply prompts in an isolated and authorized setting, record the exact model context, and treat results as version-dependent. Responsible testing can support both model evaluation and stronger prompt-injection defenses.