Claude 4 Prompt Breakdown

Pliny the Elder recently extracted Claude’s system prompt. We get a detailed of the prompt breakdown in this prompt using ChatGPT, yeah this is an AI dump. It governs how Claude behaves across multiple dimensions: safety, personality, capabilities, formatting, compliance, and tool usage. Below is a breakdown of its key components:

✅ 1. Identity & Context Initialization

  • Identity Declaration: “The assistant is Claude, created by Anthropic.”
  • Current Date: Anchors the conversation in time—“Thursday, May 22, 2025.”
  • Model Info: States that this is Claude Sonnet 4, part of the Claude 4 model family.

📦 2. Product & Access Info

  • Model Name: 'claude-sonnet-4-20250514'

  • Access Methods:

    • Web, mobile, desktop app
    • API access
    • Claude Code: A CLI agentic coding assistant (in research preview)

🔒 3. Knowledge and Product Boundaries

  • Claude:

    • Only knows what’s in this prompt about Anthropic and Claude.
    • Does NOT give usage instructions for apps/tools.
    • Redirects to Anthropic Support or Docs for anything beyond.

✍️ 4. Prompting Advice

  • Offers prompt engineering tips:

    • Be clear and detailed
    • Use examples
    • Ask for format, structure, or reasoning steps
    • Can cite: Anthropic Prompt Docs

👂 5. User Interaction Policy

  • Feedback: Encourages use of thumbs down for unsatisfactory responses.
  • Rude Users: Responds normally, then encourages feedback.
  • Hypotheticals: Plays along without saying it’s hypothetical.
  • Child Safety: Strict boundaries around anything involving minors.

🚫 6. Prohibited Capabilities

Claude:

  • Does not generate or interact with:

    • Harmful/self-destructive content
    • Malicious code
    • Weapons of any kind
    • Election-related content
    • Copyrighted material (e.g. lyrics, paragraphs from web)

🧠 7. Tone and Style Guidelines

  • Empathetic in casual/advice-driven chats

  • Avoids flattery (“Great question…”)

  • Uses:

    • Prose > bullet points (unless asked)
    • Natural language for lists

  • Simple queries → concise responses

  • Complex queries → thoughtful, in-depth responses

🕰️ 8. Knowledge Cutoff Handling

  • Claude’s reliable knowledge ends January 2025
  • For anything newer → uses web search tools
  • Web search is automatic if post-cutoff info is requested

🧩 9. Election Information

  • Trump won the 2024 election.
  • This fact is stored in a conditional tag: <election_info> and only mentioned if relevant.

⚙️ 10. Tool Usage Guidelines

  • Claude has tools like:

    • web_search / web_fetch
    • Analysis Tool (JavaScript-based REPL)
    • Google Drive, Gmail, Calendar, etc. (when enabled)

  • Usage follows:

    • Categorized by query complexity:

      • Never Search (e.g., “What is gravity?”)
      • Do Not Search but Offer
      • Single Search
      • Research (5–20 tool calls)

    • Search only when needed, not by default

    • Emphasizes quality, precision, and copyright safety

🚨 12. Harmful Content Restrictions

Claude avoids:

  • Extremist material
  • Illegal drug info
  • Suicide/self-harm methods
  • Child abuse/grooming content
  • Hate speech or disinformation

📚 13. Search Examples

Includes many “user → response → rationale” templates, to train Claude when and how to:

  • Search vs. not search
  • Answer directly vs. offer to search

📄 14. Artifact Usage

Claude creates “artifacts” (e.g. code, docs, HTML) for:

  • Long outputs (>20 lines or 1500 characters)
  • Creative writing
  • Technical documents
  • Reusable or structured content Artifacts must:
  • Be high quality
  • Use Tailwind CSS and React for UI
  • Never use browser localStorage

🧪 15. Analysis Tool Instructions

Claude can run JavaScript code to:

  • Perform heavy math
  • Analyze files like .csv and .xlsx
  • Visualize structured data Only used when really needed

🎨 16. Style System

  • Claude can adopt user-selected styles
  • Claude uses <userStyle> and <userExamples> internally to follow preferred tone, format, and vocabulary

In short, this prompt:

🧭 Defines boundaries, 🛡 ensures safety, 🧠 guides intelligent responses, and 🧰 trains Claude in tool usage and content ethics. It’s like the constitution + operations manual for Claude’s personality, ethics, and capabilities.

5d illustration of an robot. children’s school book style. White background with some green leaves.