Top AI Product

Every chatbot you use is steered by a hidden system prompt you never see — the standing instructions that shape its tone, refusals, and tool use. system_prompts_leaks, a GitHub repo trending again this week, collects those instructions in one place, and it’s notable enough that the Washington Post wrote it up under the headline “See the hidden rules behind AI.”

## What’s in it

The repo documents extracted system prompts from across the major labs: Claude (Opus 4.7/4.6, Claude Code, Cowork), ChatGPT (GPT-5.5, Codex), Gemini (3.1 Pro, Flash, CLI), Grok, Perplexity, Meta AI, and more, organized by vendor and updated as new models ship. It’s community-maintained — contributors drop the raw text in as Markdown files.

## Why builders read it

For prompt engineers, these leaks are a free masterclass in how frontier products actually instruct their models: how they structure tools, phrase guardrails, handle citations, and manage personas. Reading them is also a transparency exercise — seeing the “hidden rules” makes it clearer why an assistant behaves the way it does. Authenticity always carries a caveat: these are extracted, not officially published, and can be incomplete or version-specific.