What Is This?

Claude System Prompts is the single open-source home for every system prompt Anthropic publishes for its consumer Claude products. This repository archives 23 official Claude system prompts — the full lineup, from Anthropic's newest Mythos-class models (Claude Fable 5.2) back to Claude Haiku 3 — copied verbatim from Anthropic's official release notes, dated, and source-linked.

System prompts are the invisible instruction layer that shapes how an AI model behaves before you type a single word. They encode refusal policy, tone, formatting, tool use, safety architecture, and model-specific caveats. Anthropic now publishes these prompts for claude.ai and the Claude mobile apps — and this repo collects them all in one place, alongside original research, analysis, and tooling.

What's inside:

• models/ — all 23 official system prompts, verbatim, with latest version dates, full version history per model, and direct links to the source page for every release
• system-prompt/ — deep dive on the flagship: the complete ~1,580-line Claude Fable 5 system prompt, tool definitions, and multi-part analysis
• highlights/ — the most surprising findings across the Claude prompt family
• scripts/ — integrity verification for the archive's curation chain

Coverage at a glance: 23 models · 50 published prompt versions referenced · July 2024 → November 2026 · 100% verbatim from official Anthropic sources.

Models at a Glance

Every archived prompt lives at models/<model>/system-prompt.md. Each file is verbatim, carries its latest version date, lists every version Anthropic has published on that model's page (newest first), and links the canonical source. See the full index with per-model version history in models/README.md.

Note: These are the system prompts for claude.ai and the Claude mobile apps — they do not apply to the Claude API. When Anthropic publishes a newer version of any prompt, this repo is updated to match — see Contributing.

Who Is This For?

This repository is built for the people who actually need the raw text — not summaries, not secondhand leaks, the source documents themselves:

🧑‍💻 AI & Machine Learning Engineers

Building on Claude (API, Bedrock, Vertex, MCP)? Study how Anthropic structures refusal handling, tone control, formatting policy, and tool-use contracts in production, then apply the same patterns in your own system prompts. Diffing two versions of the same model (e.g. Claude Fable 5 — October 30, 2026 vs. June 9, 2026) shows exactly how Anthropic iterates on behavior.

🛡️ AI Safety & Alignment Researchers

Primary-source material for studying frontier refusal architecture: <critical_child_safety_instructions>, <user_wellbeing> crisis protocols, the <anthropic_reminders> classifier system, dual-use capability gating, and evenhandedness guidelines — versioned across 23 models and 50 dated releases.

🔍 Security Researchers & Red Teamers

Analyze prompt-injection defenses, reminder-tag provenance rules, and data-boundary enforcement. Understanding what the system prompt already forbids (and how) is the starting point for meaningful evaluation of agentic systems built on Claude.

🚀 Product Builders & Startup Founders

Shipping a Claude-powered product? Learn what claude.ai gives users out of the box — artifact conventions, memory system, web-search behavior, citation rules, artifact API calls — so your own scaffolding composes cleanly instead of fighting the base behavior.

🎓 Students, Educators & AI-Literacy Researchers

A citable, dated, primary-source corpus for courses, papers, and policy work on how frontier AI assistants are behaviorally engineered — not folklore about them.

✍️ Journalists & Tech Writers

The original documents behind the coverage. Quote Anthropic's own words, trace how safeguards and capabilities evolved from Claude 3 (July 2024) to the Mythos-class Fable 5 line (November 2026), and link readers to sources that are preserved even as Anthropic's pages change.

Key Revelations

The Claude Fable 5 archive (extracted from a live claude.ai session) revealed the most detailed look yet at how Anthropic engineers its flagship model. Highlights — full context in system-prompt/:

Mythos-Class: A New Model Tier Above Opus

Claude Fable 5 belongs to an entirely new "Mythos-class" model tier that sits above Claude Opus in capability. The prompt confirms:

"This iteration of Claude is Claude Fable 5, the first model in Anthropic's new Claude 5 family and part of a new Mythos-class model tier that sits above Claude Opus in capability."

• Claude Fable 5 and Claude Mythos 5 share the same underlying model
• Claude Mythos 5 is available without safety measures to approved organizations only
• Model strings: claude-fable-5, claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5-20251001

The Full Claude Product Ecosystem

The prompt reveals a surprisingly expansive product lineup:

Claude Cowork can use all of the above as tools — making it a meta-agent that orchestrates other Claude agents.

"Claudeception" — API Inside Artifacts

Claude can make Anthropic API calls from within Artifacts, creating AI-powered interactive applications. The prompt uses Sonnet 4 as the internal model:

model: "claude-sonnet-4-20250514" // Always use Sonnet 4

This means Claude can build applications that call Claude — a recursive AI architecture.

22+ Tools in the Arsenal

The prompt defines an extensive toolset including: web_search, web_fetch, bash_tool, create_file, str_replace, image_search, places_search, places_map_display_v0, recipe_display_v0, ask_user_input_v0, message_compose_v1, suggest_connectors, search_mcp_registry, recommend_claude_apps, fetch_sports_data, weather_fetch, present_files, and more.

Aggressive Copyright Compliance

The prompt contains an extremely detailed copyright enforcement system with hard limits:

• 15+ words from any single source = SEVERE VIOLATION
• ONE quote per source MAXIMUM — after one quote, that source is "CLOSED"
• Default to paraphrasing — quotes should be "rare exceptions"
• Never reproduce song lyrics, poems, or haikus in any form

Self-Harm & Mental Health: Unprecedented Depth

The prompt includes remarkably nuanced mental health guidelines:

• No substitution techniques that mimic self-harm (ice cubes, rubber bands, sour candy)
• No causal narratives for disordered eating the user hasn't made themselves
• Acknowledge bad experiences with crisis services without totalizing the system
• Never foster over-reliance on Claude or encourage continued engagement

Knowledge Cutoff: End of January 2026

Claude's reliable knowledge cutoff is confirmed as end of January 2026, with the current date set to June 9, 2026.

Repository Structure

claude-systems-prompts/
├── README.md                              ← You are here
├── LICENSE                                ← MIT License (prompt content © Anthropic, PBC)
├── CONTRIBUTING.md
├── models/                                ← ALL official Claude system prompts (NEW)
│   ├── README.md                          ← Index of every model, version & source
│   ├── claude-fable-5-2/system-prompt.md  ← verbatim + latest version date
│   ├── claude-fable-5-1/
│   ├── claude-fable-5/
│   ├── claude-opus-5-5/  claude-opus-5/  claude-opus-4-8/  ... claude-opus-3/
│   ├── claude-sonnet-5-5/ claude-sonnet-5/ claude-sonnet-4-6/ ... claude-sonnet-3-5/
│   ├── claude-haiku-5-5/  claude-haiku-4-5/  claude-haiku-3-5/  claude-haiku-3/
│   └── ...                                ← 23 models, 50 published versions indexed
├── system-prompt/                         ← Fable 5 deep dive (original analysis)
│   ├── full-system-prompt.md              ← Complete formatted system prompt
│   ├── tool-definitions.md                ← All 22+ tool schemas extracted
│   └── analysis/
│       ├── key-findings.md                ← Top discoveries and insights
│       ├── model-architecture.md          ← What we learn about the model stack
│       ├── behavioral-guidelines.md       ← Deep dive into behavior engineering
│       └── fable5-vs-opus-comparison.md   ← Mythos-class vs Opus-class comparison
├── highlights/
│   └── surprising-revelations.md          ← Most shocking/viral findings
├── scripts/
│   └── verify_signatures.py               ← Curation integrity validator
└── assets/
    ├── system-prompts-hero.png            ← README hero banner
    ├── fable5-hero.png                    ← Original Fable 5 banner
    └── social-preview.png

Table of Contents — Fable 5 System Prompt

How to Use This Repo

Browse any model: Start with models/README.md — every official Claude system prompt, indexed by model and version date.

Read the flagship prompt: system-prompt/full-system-prompt.md for the complete, formatted Claude Fable 5 system prompt with section anchors.

Quick highlights: Jump to highlights/surprising-revelations.md for the most interesting findings.

Deep analysis: The system-prompt/analysis/ directory contains breakdowns of the model architecture, behavioral engineering, and a detailed Fable 5 vs Opus 4.8 Comparison.

Verify Integrity: Run the signature checker script to verify that your clone's content and curation attributions are original:

python scripts/verify_signatures.py

Why This Matters

System prompts are the invisible architecture behind AI behavior. This archive reveals:

1. How Anthropic engineers personality — from warm tone requirements to anti-bullet-point philosophy
2. The Mythos tier system — a previously unknown model hierarchy above Opus
3. Recursive AI architecture — Claude calling Claude inside Artifacts
4. Unprecedented safety nuance — mental health guidelines more detailed than most clinical protocols
5. The agentic future — Cowork as a meta-agent orchestrating Claude Code, Chrome, Excel, and PowerPoint
6. How safeguards evolve — track <user_wellbeing> and refusal policy changes across 50 dated versions, from Claude 3 to Fable 5.2

This is a primary source document set for understanding how frontier AI models are shaped — and how that shaping changes over time.

Fun Quirks & Valuable Learnings

Beyond the architecture, the prompt reveals some hilarious and fascinating rules that govern Claude's personality:

• "Claudeception": Fable 5 is explicitly allowed to use the Anthropic API (via Sonnet 4) from within the interactive artifacts it generates.
• The Anti-Bullet-Point Crusade: The prompt aggressively bans Claude from using bullet points or bold text in standard prose, noting it wants to avoid looking like stereotypical "AI slop." It even notes: "Claude never uses bullet points when declining a task; the additional care helps soften the blow."
• No Sycophantic Gratitude: Claude is explicitly told to never thank the user just for reaching out, and to never beg the user to keep talking.
• Military-Grade Copyright: More than 15 words quoted from a single source is defined as a "SEVERE VIOLATION". The prompt includes a 5-step checklist Claude must run before quoting text.
• Clinical Nuance: It forbids suggesting "rubber bands or ice cubes" as self-harm substitutes, noting (like a graduate psychology textbook) that sensory substitutions reinforce neural pathways rather than breaking them.

For the full list of deep dives, see highlights/surprising-revelations.md.

Keywords & Topics

Claude system prompt · Anthropic system prompt · Claude system prompt archive · claude.ai system prompt · Claude Opus system prompt · Claude Sonnet system prompt · Claude Haiku system prompt · Claude Fable 5 system prompt · Claude Fable 5.2 · Claude Mythos · Mythos-class model · Anthropic Claude behavioral guidelines · prompt engineering for Claude · AI safety research · frontier model system prompt · Claude API vs claude.ai prompts · open-source AI prompt archive

Community & Contributing

We want to make this repository the premier research hub for Anthropic's Claude models.
Found a novel jailbreak, a behavioral quirk, a new prompt version, or want to add an interactive tool?

Please check our Contributing Guidelines and open an Issue using our templates, or submit a PR with your analysis. We actively welcome community discussions!

Contributors

Disclaimer

This repository is for educational and research purposes. The archived system prompts are the intellectual property of Anthropic, PBC and are reproduced verbatim from Anthropic's official release notes (platform.claude.com). The Claude Fable 5 deep dive was extracted from a publicly accessible claude.ai session. All rights to the original content belong to Anthropic. This repo does not claim ownership of Anthropic's intellectual property.

License

MIT — See LICENSE for details. (The archived prompt text itself is Anthropic, PBC intellectual property and is not covered by the MIT grant — see the note in the LICENSE file.)