Breaking Down the Leaked Fable 5 System Prompt: What's Actually Useful

A walk through what’s actually in the file, not just a promise that it’ll make your content sound human.

Anthropic pulled Claude Fable 5 on June 12, after a US export control directive cited a jailbreak that Amazon researchers had reported. It came back July 1, three weeks later, once the directive was lifted. Three weeks, start to finish.

Somewhere in that window, a jailbreak researcher published what he says is Fable 5’s complete system prompt: roughly 120,000 characters across 1,585 lines. Anthropic hasn’t confirmed the file, and neither can I. It’s been covered independently by enough outlets that I’m treating it as a useful reference. Not a verified transcript.

Most of the coverage skims it for a handful of quotable lines. Here’s what’s actually in there. And what to do with each piece.

Rules come with a reason attached

A rule on its own gets followed until it’s inconvenient. A rule with a reason attached gets followed because the reason still makes sense three paragraphs later. The prompt skips bullet points when declining a request, and says why: writing it out as prose instead, in its own words, “helps soften the blow.” Simple as that. That’s the gap between “don’t use bullets here” and a rule that survives contact with an actual draft.

Try this: next time you write a style rule, attach the because. Not “keep paragraphs short.” Try “keep paragraphs short, because this reader is scanning on a phone between meetings.” The reason is what keeps the rule from getting dropped under deadline pressure.

Warmth has a ceiling, not just a floor

Most style guides only worry about not being cold enough. This one worries about the opposite too. It tells the model not to thank someone simply for reaching out, not to invite them to keep talking, and to avoid reflective listening that risks amplifying someone’s negative emotions instead of helping them. That’s tone calibrated in both directions, not just “be nicer.”

Try this: if your brand voice guide only lists things to add, warmth, enthusiasm, personality, it’s half finished. Write down what too much of it looks like too. A support reply that over-empathizes reads as performative. A social reply that’s too upbeat next to a real complaint reads as tone-deaf. Name the ceiling, not just the floor.

Vague style words get replaced with numbers

“Be concise” means something different to everyone who reads it. The prompt mostly avoids that trap. Quotes are capped at a specific word count. Tool calls scale by task, roughly one for a single fact, several more for deeper research. Little of it is left open to interpretation.

Try this: turn your vaguest style instruction into a number. Not “keep it punchy,” but “no sentence over 25 words in the opening paragraph.” Not “cite your sources,” but “one supporting stat per claim, sourced within the last 12 months.” A number can be checked against a draft. An adjective can’t.

Boundaries get taught with contrast, not adjectives

Telling a model to be “appropriately cautious” about outdated information doesn’t tell it where the line actually sits. The prompt teaches it with a paired example instead: search for “what has this person done lately,” skip the search for “who is this person.” Same sentence shape. One variable changed. That’s the entire lesson.

Try this: next time you’re training a person or a model on a judgment call, don’t describe the boundary. That’s the move. Show two near-identical examples that land on opposite sides of it. It teaches the shape of the decision, not a definition someone can talk themselves around.

What isn’t known gets said out loud, not smoothed over

Buried in the more procedural sections is a plain instruction to state its knowledge cutoff rather than guess past it, and to search rather than assume when a question turns on something that could have changed recently: an office someone holds, a policy in effect, a company’s current leadership. No guessing. It doesn’t try to sound confident about the present. It tries to be right about it, which sometimes means saying “I don’t know, let me check.”

I ran into a version of this writing the piece before this one. “Anthropic just pulled Fable 5” was accurate the week it happened and stale by the time I was editing it, because the model came back online the same day. The fix wasn’t better writing. It was checking before publishing instead of trusting what I already believed.

Personality is a small fraction of the actual file

One breakdown of the document counted 1,585 lines total and found the “this is who you are” identity statement doesn’t show up until line 1,351. Nearly everything before it, by that account, is tool definitions, search rules, formatting logic, and safety instructions. Not tone.

Worth sitting with if your own prompt, or your own brand guide, runs 90% “sound friendly and professional” and 10% everything else. Mine used to. The ratio in the leaked file runs the other way.

The quick version

• Attach a reason to every rule, not just the rule itself.
• Set a ceiling on warmth and enthusiasm, not just a floor.
• Replace vague style words with a number that can be checked.
• Teach boundaries with contrast pairs, not adjectives.
• Say what you don’t know instead of writing around it.
• Spend your words on structure and rules, not a personality paragraph.

The prompt for running your own audit

Steve Toth’s SEO Notebook built a reusable prompt out of this same leak, one you can paste into any AI tool to audit a draft against most of the rules above. Full credit to him for the tool. Worth subscribing to ainotebook.com if this kind of breakdown is useful to you.

Audit the draft below against the rules. Do NOT rewrite it wholesale. For every spot that breaks a rule, flag it: show the original text, name the rule it breaks, and suggest a replacement in my voice. Leave anything that already follows the rules untouched - don't list it, don't change it. A sentence that's fine stays exactly as it is.

Fill in every [bracket] first, then apply the rules:

1. Voice: match the way THIS sounds, not stiffer - [paste 2-3 sentences of your own writing, OR describe the tone you want, e.g. dry and direct / warm and plain / authoritative].
2. Audience: [who's reading, e.g. busy founders / technical SEOs / first-time buyers]. Match their register and what they already know.
3. Opening: the opening sentence should already lead with the main point - a specific claim, a number, or a short anecdote. Flag any background, definitions, "in this post we'll cover," or setup that delays the substance.
4. Cut these tics: "Let's dive in," "let's unpack this," "buckle up," "picture this," "without further ado," "gone are the days of," "here's the kicker," "in conclusion," "at the end of the day," "the possibilities are endless," and "the bottom line." Don't open with "So," and don't end sections with a rhetorical question.
5. Ban these phrases and words: "in today's fast-paced digital landscape," "in the ever-evolving world of," "when it comes to," "navigate the world of," "unlock," "unleash," "harness," "leverage," "delve," "dive deep," "tapestry," "realm," "landscape," "robust," "seamless," "game-changer," "supercharge," "elevate," "empower," "crucial," "essential," "vital," "it's worth noting," "it's important to note," "that said," "moreover," "furthermore," "additionally," "notably," and "ultimately." Use a plain alternative instead.
6. Rhythm: vary sentence length - at least one sentence under [8] words per paragraph, never more than [3] long sentences in a row, and at most [1] em-dash per paragraph.
7. Format: flag formatting that fights the piece - a wall of text that should be broken up, or lists, headers, or tables padding points that would read better as plain sentences. Do not impose a word count or restructure the draft.
8. Fact pass: do not invent statistics, names, or quotes. If a number isn't in the source I gave you, flag it as unverified rather than leaving it in.

Output a single table with three columns: | Original | Rule broken | Suggested fix |. One row per flagged item; skip anything that already passes. Do not output a rewritten version of the draft - only the table.

Draft to audit:
[paste your draft]

Credit: Steve Toth, SEO Notebook (ainotebook.com).