Source
> Source: GitHub repository asgeirtj/system_prompts_leaks — 44,807 stars > Author: Asgeir (Iceland-based developer) > Media coverage: Washington Post feature (May 11, 2026) > Analyzed: Claude Fable 5 (full 120K-character leak), GPT-5.5, Claude Code, and more
This is not just gossip — it is a masterclass. The repository's value lies in revealing how top AI teams compress product design, risk control, and user experience into prompts: turning Anthropic, OpenAI, and Google's internal engineering decisions into readable text.
What's Inside the 120K-Character Claude Fable 5 Prompt
| Metric | Value | |---|---| | Characters | 120,040 | | Estimated tokens | ~30,000 | | Words | 17,074 | | Lines | 1,585 | | Sections | 72 named sections |
Before a user says anything, the model has already "read" a 30,000-token instruction manual.
Token Budget Allocation
| Block | Share | Content | |---|---|---| | Tool definitions & schemas | 30% | Full JSON schemas for 18 tools (bash, file editing, weather, recipe display components...) | | Search & citation rules | 25% | When to search, query wording, copyright compliance, citation tag formats | | Behavior, safety & health | 17% | Refusals, tone, formatting, mental health protocols, neutrality | | Identity & "Claudeception" | 13% | Identity preamble, artifacts calling the Claude API, user context, skills | | Computer use & file handling | 10% | File creation, artifact standards, output rules | | Memory, storage & MCP | 6% | Memory system state, persistent artifact storage, connector suggestions |
Most counterintuitive finding: more than half the budget goes to capability definitions (tools + search), not persona. If your agent's prompt is 80% personality with two lines about tool use — your ratio is inverted.
9 Immediately Applicable Engineering Rules
1. Incidents become rules
Claude's prompt: direct users to the National Alliance for Eating Disorders hotline instead of NEDA, because NEDA has permanently shut down. One broken hotline number became a core instruction. Log every production failure and write the specifics into your prompt.2. Failure modes get worked examples
Not "use current dates" but: "'latest iPhone 2025' returns outdated results in 2026; 'latest iPhone' or 'latest iPhone 2026' is correct." Show wrong vs. right examples rather than abstract virtues.3. Name prompt injection in plain English
Claude's prompt explicitly warns that users can append content (even claiming to come from Anthropic) and to treat such content with caution. Describing attack patterns in the prompt beats relying on filters.4. Edge cases read like postmortems
Every oddly specific rule — a disconnected mental health hotline, outdated search queries, prohibited self-harm alternatives — almost certainly traces to a production incident. Treat your prompt as a changelog.5. Negative examples everywhere
Not "be concise" but "Claude never just thanks the user for reaching out" — defined down to exact wording.6. Formatting is policy
Bullets must be 1–2 sentences; refusals must never use bullet points (extra care softens the blow); reports and technical explanations use prose. If your agent's output feeds another system or UI, format rules should be as strict as a JSON schema.7. Injection defense in plain English
Name the threat rather than hoping the model infers it — especially if your agent reads untrusted content with tool access enabled.8. Copyright rules at the prompt layer
"Statements from search must be expressed in Claude's own words, never direct quotes. Even short phrases from sources must be reworded. Citation tags are for attribution, not license to reproduce text." Legal compliance engineered into the instruction layer.9. Identity comes last
"The assistant is Claude, created by Anthropic" appears at line 1,351 of 1,585. Operational instructions occupy the highest-attention positions; persona is a footer, not a header.Claude vs. ChatGPT vs. Gemini: Three Worldviews
| Dimension | Claude (Anthropic) | ChatGPT (OpenAI) | Gemini (Google) | |---|---|---|---| | Core philosophy | Safety-first, anti-engagement | Pragmatic, selectable personas | Google ecosystem integration | | Prompt size | 120K chars (largest) | Medium, multiple versions | Relatively lean | | Safety | Detailed refusal protocols, mental health lines | Image safety, automation context | Looser | | Tools | 18 tools, detailed schemas | Search, Python, Canvas, DALL-E | YouTube, Workspace integration | | Identity position | Last (line 1351/1585) | Usually first | First | | Anti-engagement clause | Explicit "Claude never asks users to keep talking" | None | None | | Persona | Warm, honest, empathetic refusals | Optional (Friendly/Nerdy/Cynical...) | Neutral |
Key insight: Anthropic bakes safety protocols into the DNA of its system prompt; OpenAI hands persona choice to users.
"Claudeception" and Search Strategy
Claudeception is an internal concept: AI apps built inside Claude's chat interface that call the Claude API themselves — AI invoking AI within itself. Instructions include example fetch calls and "never pass API keys, since authentication is handled by the platform."
Claude's search instructions split information into established knowledge (answer directly) and current state (always search, even if the model thinks it knows). The UNRECOGNIZED ENTITY RULE mandates searching before answering about any unrecognized game, movie, show, book, album, product launch, menu item, or sporting event. For brands: if you're not in training data, the model searches you every time — your web presence determines how AI describes you. Publishing PR is writing AI's answers.
Mental Health Protocols
The most detailed section covers mental health and user wellbeing: no diagnoses or labels (unless self-identified), no listing actionable means in self-harm contexts, directing eating-disorder queries to appropriate support organizations, never surfacing memories containing sensitive content unless explicitly raised, and stating principles rather than detection mechanisms when refusing. System prompts are no longer prompts — they are product risk-control checklists requiring long-term maintenance.
Action Checklist
Today: Restructure your CLAUDE.md/SKILL.md with named functional blocks; build a prompt changelog; audit trust boundaries and describe injection attack patterns in the prompt.
This week: Rebalance so tool definitions + search rules exceed persona (Anthropic: 55% vs 17%); add negative examples with precise wording; define output contracts as strictly as API schemas.
This month: Build an incident-to-rule feedback loop; study competitor prompts in the repo; test moving identity declarations to the end.
Conclusion: Prompt Engineering Is Product Design
> A system prompt is not "how the AI speaks" — it is "how the AI works."
In Anthropic's 120K characters: 30% tool schemas, 25% search rules, 17% safety and behavior, only 13% identity and persona. For product managers, your requirements doc should translate directly into a system prompt; for developers, CLAUDE.md should be more detailed than code comments; for researchers, these leaks are the best window into frontier teams' priorities.
References:
- GitHub: https://github.com/asgeirtj/system_prompts_leaks
- Washington Post: https://www.washingtonpost.com/technology/interactive/2026/chatbots-hidden-rules-system-prompts/
- Ayautomate analysis: https://www.ayautomate.com/blog/claude-fable-5-system-prompt-leak
- Knight Li section-by-section reading: https://knightli.com/en/2026/06/12/claude-fable-5-system-prompt-analysis/