Inside the Leaked Claude Fable 5 System Prompt

On June 9, 2026, Anthropic released Claude Fable 5, its most capable widely available model. Within about a day, the jailbreak researcher known as Pliny the Liberator published a roughly 120,000-character file claiming to be the model's full system prompt — the hidden instructions that sit in front of every conversation on Claude.ai. The file spread fast through the CL4R1T4S GitHub repository and X. It is worth reading, but only with a clear caveat up front: Anthropic has not confirmed it, and prompt-extraction methods can introduce gaps or hallucinated passages. Treat it as an unofficial reconstruction, not gospel.

What actually leaked

The first thing to get right is scope. By one widely shared analysis the circulating file runs to 120,040 characters across 1,585 lines, organized into 72 named sections — roughly 30,000 tokens consumed before a user types anything. Crucially, it describes the prompt for the Claude.ai consumer chat interface, not the model in general. That distinction matters for builders: on the Claude API there is no system prompt by default, and Claude Code ships its own, very different instructions. So even if the extraction is accurate, it is not "how Fable 5 thinks" — it is how one product wraps the model.

It reads like a product spec, not a personality

The popular mental image of a system prompt is a paragraph of vibes — "be helpful, be honest." The leaked file is the opposite. The same analysis breaks its token budget down as roughly 30% tool definitions and schemas, 25% search and citation rules, 17% behavior/safety/wellbeing, 13% identity and artifacts, 10% computer use and file handling, and 6% memory and MCP apps. Capability plumbing dominates. The famous identity line — "The assistant is Claude, created by Anthropic" — reportedly does not appear until line 1,351 of 1,585, about 85% of the way through. Whoever wrote it clearly considered tool behavior and search hygiene more load-bearing than self-description.

Negative examples beat lofty virtues

The most useful pattern for anyone who writes prompts is how the file phrases its rules. It rarely settles for an abstraction like "be concise." Instead it spells out the exact thing not to say. Two examples that circulated widely: Claude "never thanks the person merely for reaching out," and "never use bullet points when declining a task." This is a real prompt-engineering lesson, independent of whether every line is authentic — concrete negative examples with specific wording tend to steer a model more reliably than vague positive traits. If you want a behavior gone, naming the precise failure usually beats describing the ideal.

The edge cases read like incident postmortems

Several of the oddly specific lines look less like design and more like scar tissue from things that went wrong in production. The file reportedly instructs Claude to point users to "the National Alliance for Eating Disorders helpline instead of NEDA, because NEDA has been permanently disconnected" — a fix for a real-world helpline that shut down. Another guards against stale search results with the example "'latest iPhone 2025' when the year is 2026 returns stale results." It also enumerates sensitive self-harm topics the model must never assist with. Each hyper-specific clause almost certainly maps to an incident that was patched directly into the prompt — which is why production system prompts grow long and strange over time.

Why Claude "writes in prose now"

One detail resolved a question many users had been asking: why recent Claude output leans on paragraphs instead of bullet lists and headers. The leaked file reportedly tells the model to avoid bullet points, headers, and lists unless the user asks for them. This direction is consistent with Anthropic's own public prompting guidance for Fable 5, which repeatedly stresses readable prose, leading with the outcome, and dropping "arrow-chain shorthand" and hyphen-stacked compounds in user-facing messages. So even without the leak, the prose shift is documented behavior — the file just shows the blunt instruction behind it.

The jailbreak claim — and Anthropic's pushback

The same researcher also claimed a jailbreak of Fable 5, and Anthropic disputed it directly. According to the company, the demonstrated technique was not a genuine bypass but "coaxing the model to continue responding despite its conversational refusals" — a known limitation, not a defeat of core safeguards. Anthropic stressed that its strongest protections run as independent classifier systems separate from the model, so getting past a conversational refusal does not switch them off. On review, it said some shared outputs were not generated by Fable 5 at all, and the genuine ones contained only information already public, "offering no meaningful uplift for real-world harm." The takeaway: extracting a system prompt is not the same as breaking safety, and the two claims should not be conflated.

What builders should take from it

A few practical lessons survive the authenticity question. First, the consumer prompt is not your prompt — on the API you start from a blank slate, so don't assume Claude.ai's tool rules or tone apply to your integration. Second, Fable 5 itself has hard behavioral specifics worth knowing from Anthropic's docs: a 1M-token context window, adaptive thinking only (raw chain-of-thought is never returned), an effort control instead of thinking budgets, and a refusal stop reason when its safety classifiers (covering offensive cyber, biology/life-sciences, and reasoning-extraction) decline a request — with fallback to Claude Opus 4.8. That last point connects back to the leak: prompts that tell the model to echo or transcribe its own reasoning can trigger the reasoning-extraction classifier, so "show me your system prompt"-style instructions are exactly what the model is trained to refuse.

Bottom line

The leaked file is best read as a user manual for how Claude.ai is steered, not a smoking gun — and not a confirmed document. Its real value is craft: front-load tool and search behavior, encode rules as precise negative examples, and treat each weird edge case as a logged incident rather than noise. Until Anthropic publishes an official version, verify any specific line before you build on it. Expect more of these extractions with every frontier launch — and expect the gap between "leaked the prompt" and "broke the model" to keep getting blurred in the headlines.

FAQ

Is the leaked Claude Fable 5 system prompt official? No. It is an unofficial extraction published by an independent researcher. Anthropic has not confirmed it, and extraction can introduce gaps or fabricated lines.

Does it apply to the API or Claude Code? No. It describes the Claude.ai consumer chat interface. The API has no default system prompt, and Claude Code ships its own separate instructions.

Was Claude Fable 5 actually jailbroken? Anthropic says no. It characterized the technique as coaxing past conversational refusals, not bypassing its independent safety classifiers, and said some shared outputs weren't even from Fable 5.

How large is the system prompt? By one analysis, about 120,040 characters, 1,585 lines, and roughly 30,000 tokens — spent before any user message.

Sources and further reading