Claude Prompt Leak
Anthropic’s leaked Claude 4 system-prompt shows—line by line—how the model is instructed when to fire a live web search and whether to surface a clickable citation.
Why is this important?
When you're optimizing your content, you should start with the content that is already cited by Ai. This content will be picked up immidiately by Ai, just as in the above screenshot, Expatsian is used to list the price of bread. Because that page is used frequently to answer questions, if changes are made to it, citations by ChatGPT will include the new content almost immidiately.
At RankBee.ai we found that 25% of Ai prompts trigger ChatGPT Web Search. So when optimizing your content using RankBee.ai, start first, by optimising pages with citations.
Claude Leak Summary:
Claude is instructed to link when the target page does something the model cannot trivially paraphrase.
- Interactive tools: The output is personalised and can’t be pre-baked into model weights. Think Mortgage payoff calculator, tyre-size matcher, build-your-own-PC wizard
- Continuously updated datasets: Real-time airfare trackers, commodity price dashboards, election night tallies.
- Unique user reviews & ratings: Subjective, rapidly evolving, legally safer to link than to scrape. Think Verified-owner EV range logs, crowdsourced campground ratings
- Regional / niche reports: Best off-grid climbing spots in Kalymnos this summer.
- Expert commentary & analysis: Example: In-depth analysis of Google’s March 2025 Core Update, complete with SERP volatility charts.
Citations:
Hanns Kronenberg
