Every Fable 5 chat runs on a ~25,000-word operating manual — persona, ~30 tools, and pages of safety rules. A copy surfaced in a public "system prompt leaks" repo. Here's what it reveals about how the model actually behaves, and what it teaches you about writing your own. It's Anthropic's document, so we quote only short excerpts and link the full source.
The "character" everyone talks about is a sliver of the document. The bulk is tools, safety, and formatting discipline — roughly 25,000 words, with a 190,000-token budget declared in the very first lines.
Around 30 tools are defined right in the prompt: a bash shell, web search and fetch, persistent memory, search over your past chats, file creation, artifacts, data-viz, and third-party connectors. Behavior and capability live in one document.
There's no "just be safe." Instead, concrete, separately-scoped rules for child safety, weapons, drugs, malware, self-harm, disordered eating, and medical/legal advice — each with hard do's and don'ts.
On Anthropic's own products, it's instructed to search the live docs before answering — because the prompt itself goes stale. A model told, in writing, not to trust what it thinks it knows.
The takeaway isn't the vibe — it's the discipline. The most capable model on earth ships with more pages of what not to do than of who to be.
Multiple passages tell it to avoid bullets, headers, and heavy bold, and to write prose — even that a list should read as "some things include: x, y, and z" inside a sentence. The model famous for bullet soup was explicitly told not to.
The most interesting rules are meta. On sensitive requests: if it notices itself reframing a request to make it acceptable, that noticing is the cue to refuse — not a reason to proceed.
When it declines, it's told to state the principle, not the detection mechanics — because explaining exactly where the line sits teaches people how to reframe around it. It applies that even to its own reasoning.
Small behavioral tells: at most one question per response, and never bullet points when declining a task — because the extra care "softens the blow."
"Be warm" does nothing. Anthropic spells out enforceable behaviors: warm but willing to push back, one question max, minimal formatting, no cursing unless matched. Write your prompts as behaviors, not adjectives.
Even the world's most-polished prompt spends paragraphs suppressing lists and bold. If your outputs are bullet soup, a one-liner won't fix it — you have to over-specify the format you actually want.
Vague "be safe" loses. What works: concrete per-domain do's and don'ts, plus meta-rules about the model's own reasoning — the "if you're reframing it to make it OK, refuse" pattern.
The single most portable idea here: instruct your agent to go check the live source instead of trusting stale internal knowledge. Anthropic does exactly this for its own products — so should you.
We kept the quotes short and did the analysis in our own words. The complete system prompt lives in a public repository — read the original there and judge for yourself.