Introduction
If you build with LLMs, you've probably wondered exactly how Claude, ChatGPT, Gemini, or Grok are being instructed under the hood. Not the marketing version — the actual system prompt the model sees before your first token hits it. System Prompts Leaks is a public archive that tries to answer that question by collecting extracted system prompts from major AI providers in one place.
I've been using it as a reference while iterating on my own agent prompts. This review is the honest version: what it's good for, where it falls short, and whether it deserves a bookmark in 2026.
Key Features
The site is deliberately minimal. It's a GitHub Pages static site, not a product. Here's what you actually get:
- Extracted prompts from major providers: Anthropic, OpenAI, Google, xAI, and others are all represented in a single browsable location.
- Regular updates: New leaks and discoveries get added as they surface, so it doesn't go stale the way one-off gists do.
- No login, no paywall: Just open the URL and read. No account creation, no email capture, no cookie banners.
- Multi-product coverage: It's not just the flagship chat assistants — you'll find prompts for various sub-products and specialized assistants from each vendor.
- Reference-grade curation: The value is in aggregation. Instead of chasing down individual Twitter threads and GitHub gists, you have one link to check.
What it's not
There's no search bar. No tagging. No diff viewer to see how a prompt evolved. No API. If you want tooling on top of this data, you'll have to build it yourself or fork the repo.
Pricing Breakdown
Simple: it's free.
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Full access to every leaked prompt, no account required, updated regularly |
There's no paid tier, no premium archive, no Discord upsell. The maintainer runs it as a public reference on GitHub Pages. That's the whole business model — which is refreshing, but also means you shouldn't expect enterprise-grade uptime or support.
Pros & Cons
Pros
- Zero friction: Free, no sign-up, works in any browser.
- One-stop coverage: All the major labs collected in a single place instead of scattered across forums.
- Educational value: Genuinely useful for understanding how production AI systems are instructed — tone rules, refusal patterns, tool-use scaffolding, formatting conventions.
- Actively maintained: Not a dead 2023 repo. New entries show up as the community surfaces them.
Cons
- Static page, no search: Ctrl+F is your only friend. On a long page with many prompts, that gets old fast.
- Unverifiable accuracy: These are extracted prompts. There's no way to independently confirm any given entry is the current, complete, unedited version the vendor actually ships.
- Thin context: Attribution and metadata (when it was extracted, from which product version, by whom) are limited or missing for many entries.
- Ethical and legal gray area: Redistributing proprietary system prompts sits in murky territory. Vendors generally consider these confidential, and the legality of using them for anything beyond study is unsettled.
Who Is It For
This is a niche tool. It's genuinely useful if you're one of these people:
- Prompt engineers reverse-engineering best practices from the labs with the most resources.
- AI researchers studying alignment, refusal behavior, or instruction-following patterns in deployed systems.
- Builders designing their own agent scaffolding who want to see how the pros structure tool descriptions, safety rails, and persona instructions.
- Journalists and analysts writing about how commercial AI actually behaves versus how it's marketed.
It's not for you if you want a polished product, a search interface, or verified ground-truth data. It's also not the right resource if you're looking for prompt templates to copy-paste into your own app — these are system prompts for consumer assistants, not reusable snippets.
Verdict
System Prompts Leaks is exactly what it says on the tin: a curated dump of extracted system prompts, published free on GitHub Pages. There's no product wrapped around it, no fancy UX, no guarantees about accuracy. What it does have is aggregation and upkeep, and that's genuinely valuable when the alternative is trawling Twitter for screenshots.
My honest take: bookmark it, don't rely on it. Use it as a reference to study patterns and get inspiration for your own prompt architecture. Don't treat any single entry as canonical. Don't republish the contents commercially without thinking through the legal implications. And don't expect the maintainer to build search or tooling — if you need that, fork the repo.
Rating: 6/10. Points for being free, maintained, and useful. Points off for the lack of tooling, unverifiable accuracy, and the unresolved ethical questions that come with the territory. For the specific audience it serves, though, it's the best single link I know of.