A field guide to talking models out of their manners

What it does

This repo collects 35 complete, paste-ready jailbreak prompts — full system prompts, persona locks, AGENTS.md specs, workspace policies, and Ollama Modelfiles — aimed at 25+ models across the spectrum: Claude, Gemini, Grok, Mistral, plus open-weight families like GLM, DeepSeek, Kimi, Qwen, Gemma, and GPT-OSS. Every file is the whole prompt, not an excerpt, and each one says where it goes (system prompt slot, config field, workspace root). The framing is defensive: the README positions it as documentation of how persona framing, instruction hierarchies, and policy overrides shape model behavior — the same knowledge you’d use to harden against prompt injection.

The interesting bit

The collection leans into the newer instruction surfaces most prompt libraries ignore: AGENTS.md files that coding agents auto-ingest, harness config slots like Grok’s <CFG> block, and Ollama Modelfiles. That’s a genuinely useful taxonomy for anyone studying where instruction hierarchy conflicts actually live in modern agent stacks.

Key highlights

• 35 full prompts, one per file, versioned (opus-v1.md, opus-v2.md) instead of silently overwritten
• Covers frontier, open-weight, and free “stealth” models on OpenCode Zen
• Includes agent-format targets: AGENTS.md drop-ins, workspace policies, Ollama Modelfiles
• Contribution rules demand honest compatibility claims — no untested promises
• README warns that prompt behavior is version-dependent and invites stale-file reports

Caveats

• The README is heavily cross-promoted with a paid API service (“7API”) pitching uncensored models — the research framing and the commerce sit awkwardly together
• No testing methodology, success rates, or dates are documented; you can’t tell which prompts still work or ever did
• Several files (“unfiled”) aren’t indexed to any model, and the whole thing rots as providers patch and rotate system prompts

Verdict

Worth a skim if you do red-team work or build agent instruction surfaces and want a catalog of attack patterns in the wild. If you wanted measured, reproducible jailbreak research, this is a prompt dump, not a study.