A repository has captured system prompts from various AI models, revealing significant differences between the publicly released versions and those used internally by the models. For instance, Anthropic's Claude has a published system prompt of 4,269 words, while the captured version used in production is 56,148 words. This discrepancy highlights that published prompts often omit operational details, tool definitions, and specific instructions. The analysis of these prompts indicates a substantial focus on refining the AI's tone and voice, as well as more granular refusal categories than publicly stated safety policies. AI
IMPACT Reveals how AI companies focus on tone and granular safety controls, offering insights for prompt engineering.
RANK_REASON Analysis of system prompts for AI models, not a new model release or core research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
