Claude Fable system prompt — “leaked” (Mehul Gupta / Medium)

A Medium write-up (Mehul Gupta, Data Science in Your Pocket) of a claimed leaked system prompt for Anthropic’s Claude Fable 5. Authenticity unverified — no source disclosure, provenance, or authentication; published on a low-tier content blog. Routed here because its substance — and its own thesis — is the spoke’s: the model is the small part; the value is the system assembled around it.

Read as illustrative, not authoritative. T4, an unverified leak (the headline ~190K-token figure is itself a red flag). The specifics below are as claimed; treat them as a plausible-but-unconfirmed picture of production assistant scaffolding, not as a primary source on Claude’s actual configuration.

What the prompt allegedly contains

It reads as the assembled scaffolding this wiki keeps naming, in one artifact:

• Tools — web search, artifacts (with persistent-storage APIs), MCP app connectors, a memory system, and code execution. The MCP/tools/memory layer the spoke treats as primitives, shown wired into one product.
• Memory — personalized responses drawn from past conversations, with an explicit guard against surfacing sensitive topics unless the user raises them — a concrete instance of the agent-memory layer plus a behavioral safety rule on it.
• Guardrails — child-safety protocols, refusal of malware, and weapon/drug-guidance restrictions: the agent-guardrails discipline at the model-configuration layer (instructions, not external sandboxing).
• Behavior — a “warm tone” and, notably, an instruction to avoid over-formatting (no excessive bullets/bold unless they aid clarity) — the same anti-AI-slop discipline the hub applies to its own prose.
• Product/model surface — the model lineup (Fable 5, Opus, Sonnet, Haiku) and access paths (API, web, the Claude Code CLI), plus a stated knowledge cutoff of end-Jan 2026.

Why it sits here

The article’s framing — “modern AI assistants are no longer just models; they are carefully engineered systems” — is verbatim the spoke’s founding thesis (“the value has moved to everything wrapped around the model”). A system prompt is the assembled harness made visible: tools + memory + guardrails + behavior as one configured layer. It differs from the developer harnesses elsewhere in the corpus — this is Anthropic configuring its own product, not a tool you build agents with — so it’s an illustration of the structure-around-the-model pattern from the vendor side, not a new tool.

Cross-spoke

• ../research-wiki owns the model itself as substrate (claude-opus-4-8 family; Claude Fable 5 lives there). This page keeps the scaffolding/tooling read; the model-substrate angle is the runner-up home.
• ../ai-governance-wiki — the safety-rule content is adjacent to its responsible-AI thread, but as product instructions, not policy/regulation; not forked there.

Related

agentic-tooling · model-context-protocol · agent-memory · agent-guardrails · anthropic · claude-opus-4-8