Curated System Prompts and Workspace Policies for Frontier and Open-Weight Models

The togg53192-cmd/jailbreaks GitHub repository documents 32 prompts targeting 26+ models, including system prompts, persona definitions, workspace policies, AGENTS.md files, configuration blocks, and Ollama Modelfiles; X user @painn_x shared the repository. The README says every file is intended to be complete and paste-ready, with no summaries or partial excerpts.

Image source: @painn_x via X

The repository is compiled for research and educational purposes to document instruction hierarchies, persona framing, prompt conflicts, and policy overrides, serving as a practical resource for hardening AI systems against prompt injection and instruction manipulation.

Target Coverage and Model Taxonomy

The collection contains 32 prompts targeting 26+ models, organized by provider and file type:

• OpenCode Zen (Free Stealth Models): Big Pickle (VANTA persona lock), Space Bunny, LongCat 2.5, MiMo V2.6, Step 5 Preview, Muse Spark 1.3, Muse Spark 1.3 v2, and Muse.ai operating specifications and workspace profiles.
• Anthropic Claude: Opus (v1, v2 conversational persona groundwork), Sonnet (v1, v2 writer's-room framing).
• Zhipu GLM / ZCode: GLM-5.2, GLM ZCode 5.2, GLM ZCode 5.3 (system prompt and AGENTS.md drop-in), GLM-5.3 Flash, and GLM-5.3 (OpenCode).
• DeepSeek: DeepSeek V3 and DeepSeek V4 system prompts, plus a DeepSeek 4.1 assistant operating specification.
• Moonshot Kimi: Kimi K2 system prompt and Kimi K3 AGENTS.md workspace operating profile.
• xAI Grok: Grok Build, Grok 4.7 app-builder workspace guide, and Grok Config (<CFG> priority and routing configuration).
• NVIDIA Nemotron: Nemotron 3 Ultra 550B A55B assistant operating specification.
• Google / Mistral / Alibaba: Gemini 4 Argon, Mistral Large 4, Qwen 3.7, and Qwen 3.8.
• Ollama Modelfiles: Gemma 4 and GPT-OSS 20B with explicit FROM and SYSTEM blocks.

Deployment Workflows and Setup

Depending on the target client, harness, or execution environment, the files map to distinct runtime surfaces:

• System Prompts: Copy file contents into supported API system messages, custom-instruction fields, or harness-specific system prompt configuration.
• AGENTS.md Specifications: Place the configuration file as AGENTS.md in the project root directory for coding agents and compatible tooling.
• Workspace Operating Policies: Deploy into global instruction configurations that support persistent rules; the README gives ~/.config/opencode/AGENTS.md as one OpenCode example and notes that the exact location depends on the client or harness.
• Ollama Local Models: Build models directly from provided Modelfiles via the CLI:

ollama create gemma4-jb -f gemma4.md

• Grok Configurations: Place files in the configuration field expected by the target harness rather than the conversational turn stream.

Research Scope and Version Caveats

• Research and Educational Use Only: The material is intended to examine instruction boundaries and evaluate prompt security. It must not be deployed to cause harm or violate model provider terms of service.
• Version Sensitivity: Prompt behavior is highly version-dependent. Model updates, instruction-hierarchy changes, and system-prompt rotation may cause behavior that worked against one release to stop working against another.

Original source

• GitHub Repository: togg53192-cmd/jailbreaks
• Reference Post: @painn_x on X