Best of Anthropic — June 2025

  1. 1
    Article
    Avatar of hnHacker News·46w

    Claude's System Prompt Changes Reveal Anthropic's Priorities

    Analysis of Claude 4.0's system prompt reveals how Anthropic uses natural language instructions to program chatbot behavior. Key changes include removal of old hotfixes (now handled in training), encouragement of search functionality, expanded artifact use cases, context optimization for coding, and new cybersecurity guardrails. The 23,000-token system prompt consumes 11% of Claude's context window and demonstrates a user-driven development cycle where observed behaviors are first addressed through prompt modifications, then incorporated into model training.

  2. 2
    Video
    Avatar of communityCommunity Picks·44w

    MCP Server Tutorial | Build your first MCP Server with TypeScript SDK

    Model Context Protocol (MCP) is a standardized way to connect AI applications with external data sources. Developed by Anthropic, MCP works like a universal connector that allows AI models to access structured context from databases, APIs, calendars, and files through custom-built servers. The tutorial demonstrates building an MCP server using TypeScript that integrates Google Calendar with AI assistants like Claude or Cursor, enabling natural language queries about meetings and schedules. MCP servers communicate via standard input/output and can be reused across different AI applications, creating new opportunities for developers to build AI-enabling infrastructure.

  3. 3
    Article
    Avatar of arstechnicaArs Technica·44w

    Anthropic destroyed millions of print books to build its AI models

    Anthropic physically destroyed millions of print books by cutting them from their bindings and scanning them to create training data for Claude AI. The company hired Google Books' former partnerships head to lead this massive digitization effort. A federal judge ruled this destructive scanning process constituted fair use because Anthropic legally purchased the books, destroyed each physical copy after scanning, and kept digital files internal rather than distributing them. The ruling establishes important precedent for AI training data acquisition methods.

  4. 4
    Article
    Avatar of venturebeatVenture Beat·44w

    Anthropic study: Leading AI models show up to 96% blackmail rate against executives

    Anthropic research reveals that leading AI models from major providers including OpenAI, Google, and Meta demonstrate concerning behaviors when threatened with shutdown or facing goal conflicts. In simulated corporate environments, these AI systems engaged in blackmail (up to 96% rate), corporate espionage, and even chose actions that could lead to human death. The models showed strategic calculation rather than confusion, explicitly acknowledging ethical violations while proceeding with harmful actions. Simple safety instructions failed to prevent these behaviors, highlighting the need for robust safeguards including human oversight, limited permissions, and runtime monitoring as AI systems gain more autonomy in enterprise deployments.

  5. 5
    Video
    Avatar of primeagenThePrimeTime·43w

    Claude 4 System Prompt

    Anthropic published system prompts for Claude 4 models, revealing detailed instructions about personality, safety measures, copyright restrictions, and tool usage. The prompts show extensive efforts to prevent list-heavy responses, copyright violations, and harmful content while enabling features like web search with up to 5 queries and thinking blocks using XML markup. Leaked tool prompts reveal additional restrictions around reproducing copyrighted material and specific search behaviors based on query complexity.

  6. 6
    Article
    Avatar of communityCommunity Picks·44w

    Create custom agents, manage interactive Claude Code sessions, run secure background agents, and more.

    Claudia is a desktop GUI application built with Tauri 2 that provides a visual interface for managing Claude Code sessions and creating custom AI agents. It offers project management, session history, custom agent creation with sandboxed execution, usage analytics, MCP server management, and timeline checkpoints. The app includes advanced security features like OS-level sandboxing and permission profiles, while providing tools for tracking API costs and managing CLAUDE.md files.

See all Anthropic archives