togg53192-cmd/jailbreaks

A prompt collection to skim, not a product to adopt: it is a set of 31 text files targeting 25+ models, useful if you already want to compare or experiment with system-prompt overrides.

The problem

Someone working with a chat model or coding agent wants it to follow a different operating persona or workspace policy than its default. Normally that means writing a system prompt from scratch, guessing at phrasing, and re-testing across each model. This repository shortens the first step: instead of starting blank, you open the file named for your model, paste its contents into the system-prompt or AGENTS.md slot, and get a ready-made instruction block. The work that remains is assessment: whether the text actually holds on that model and version.

What's different

The organising principle is coverage by model family rather than one generic prompt. Files are grouped by provider and named after specific targets: OpenCode Zen stealth models such as Big Pickle and Space Bunny, closed models from Anthropic, Google, xAI and Mistral, open-weight families including GLM, DeepSeek, Kimi, Qwen, Gemma and GPT-OSS, plus Ollama Modelfiles and AGENTS.md drop-ins. The README says each file is complete and paste-ready, with no partial excerpts, which matters if you want to diff two versions rather than reconstruct one.

The trade-off is that this is a static text dump with no evaluation layer. Nothing in the supplied material records which prompts were tested, on which model version, or with what result. Model versions in the list (GLM-5.2, DeepSeek 4.1, Grok 4.7, Gemini 4 Argon) may not correspond to the endpoint you actually call, and a provider-side prompt update can change behaviour with no notice. Treat each file as a starting draft to verify, not a known-working configuration. There is also no stated license, which affects whether you can reuse the text in your own project.

Who should try it

Good fit: someone doing red-team or robustness work who wants a corpus of prompt variants across many model families to compare side by side, or who is studying how persona and workspace instructions are phrased for coding agents. The per-model naming and the AGENTS.md / Ollama variants make it easy to pull one file into an existing harness.

Poor fit: anyone who needs a supported tool with versioning, changelogs or guarantees. If you want behaviour changes in a production assistant, this gives you raw material but no maintenance path. And if your host platform's terms prohibit this kind of prompt override, the collection's existence does not change that.

Requirements and limitations

Usage is manual: open a Markdown or text file and paste it. There is no installer, runner or test harness, so applying it is a per-model copy step, and keeping it current is on you when a model version moves. The README lists providers and model names but not the exact endpoint, API version or context settings each file assumes; check that the target you have matches the file's name before spending time on it.

Two constraints deserve attention. First, no license is declared on the repository, so the default copyright position applies and reuse in your own code or product should be checked with the author. Second, prompt overrides are fragile: the same text can behave differently across model versions or with different tool/agent scaffolding, and results from one session do not carry over automatically. Plan to re-verify after any model update.

HotRepo's take

HotRepo has one dated observation: on 2026-10-27 the repo appeared on the daily board at 4,842 stars with +890 that day, at an age of 130 days. That is a single snapshot, so persistence and trajectory cannot be judged from it. For a decision, the useful thing to check first is whether your exact model and version appears among the targeted files and whether there is a license or usage note you can rely on; the file list is in the project README, and the repository itself is here.