Every chatbot ships with a hidden layer of instructions its makers rarely publish — the rules that quietly govern tone, refusals, formatting, and tool use. This repo treats those system prompts as primary documents worth preserving: it gathers them across vendors and dates each capture, so you can watch how a model's invisible rulebook drifts from one release to the next instead of guessing at why behavior changed.
What Sets It Apart
- Cross-vendor in one place. Anthropic, OpenAI, Google, xAI, Microsoft, Meta, Mistral, Perplexity, Qwen and others sit side by side, so you can compare how rival labs phrase the same guardrails rather than hunting across scattered gists and screenshots.
- Versioned, not just dumped. Prompts are filed by product and version with update dates, so a diff between, say, a newer and older Claude tells you what the vendor actually changed — useful signal, not just today's snapshot.
- Raw extractions over paraphrase. Entries preserve the exact wording, ordering, and tool definitions, which is what prompt engineers need to reason about why outputs look the way they do.
- Kept current. It is updated regularly and backed by a large community (40k+ stars, thousands of forks), so new models tend to show up soon after release.
Who Should Use It
Great fit if you reverse-engineer model behavior, design your own production system prompts, or audit how vendors handle safety and refusals — concrete examples beat speculation. Look elsewhere if you need official, guaranteed-current specs: these are community-extracted, may lag silent vendor updates, and reflect what someone managed to capture, not anything a vendor formally endorses.
