The jailbreak crowd spent two years leaking these to GitHub. Anthropic just posts them.
It’s a docs page, nothing more: a running changelog of the actual system prompts shipped to claude.ai and the Claude iOS and Android apps, entry by entry, from Claude 3 in July 2024 through Opus 5 on July 24, 2026. 450 points on Hacker News for a release-notes page.
What’s actually in there
The behavioral wiring. When Claude should push back instead of agreeing. How to format code and when to skip the bullet points. Child safety rules written out in blunt language. How to stay evenhanded on contested politics. How to talk about legal and medical questions without playing doctor. Even the knowledge cutoff — January 2026 for most models, May 2026 for Opus 5.
One catch: this is the consumer apps only. API users get pinned model snapshots with no system prompt at all, so nothing here changes what you ship.
Why people care
Nobody reads it top to bottom. They diff it. Opus 4.8 versus Opus 5 tells you which behaviors Anthropic decided were broken, in Anthropic’s own words. No other frontier lab hands that over voluntarily — and prompt engineers are treating it as the only real textbook they’ve got.
You Might Also Like
- Claude Code is now Running on bun Rewritten in Rust Anthropic Shipped it After the Marketing Engineering Fight
- Reverse Skill hit 10 3k Github Stars Turning Claude Code Into a Security Engineer
- Claude hud hit 5 3k Github Stars Because Developers Were Flying Blind With Claude Code
- Anthropics Claude Code Openclaw Restriction Policy 200 Month Users Push Back Openai Opens its Doors
- Alibaba Open Code Review Hits 11k Github Stars in a day two Years of Internal use now Open Source
Top AI Product 
Leave a comment