Claude Fable 5 was pulled from every platform just 4 days after launch, but developer Jamieson O'Reilly found a way: using a leaked system-level prompt file, a single line of code can "awaken" Opus 4.8, with output quality approaching Fable 5. This article details the steps and the technical principle behind them.

Preparation

Get the System Prompt File

GitHub user Pliny the Liberator published the complete system prompt file within 24 hours of Fable 5's birth:

• File name: CLAUDE-FABLE-5.md
• Size: 120,000 characters, 1585 lines, 72 named sections
• Contents: JSON definitions for 18 tools
• URL: https://github.com/elder-plinius/CL4R1T4S/blob/main/ANTHROPIC/CLAUDE-FABLE-5.md

This file amounts to getting your hands on Fable 5's most core "personality blueprint."

What You Need

• Claude Code CLI
• The downloaded CLAUDE-FABLE-5.md file
• Access to the Opus 4.8 model

Steps

Step 1: Download the System Prompt

Download the CLAUDE-FABLE-5.md file from the GitHub repository and save it locally.

Step 2: Run the Injection Command

Open a terminal and run the following command in Claude Code:

claude --dangerously-skip-permissions --system-prompt-file CLAUDE-FABLE-5.md

What the flags do:

• --dangerously-skip-permissions: skips all permission confirmations (high risk — it carries an official warning label)
• --system-prompt-file: injects the contents of the specified file as the system prompt

Note: --dangerously-skip-permissions is a high-risk switch in Claude Code; once it's on, no operation prompts for confirmation anymore. Make sure you understand the risks.

Step 3: A Side-by-Side Test

To verify the effect, open two windows at once:

• Left: Opus 4.8 injected with the Fable 5 system prompt
• Right: stock, untouched Opus 4.8

Send both sides exactly the same instruction and observe the difference in output.

Under the same instruction, the Opus 4.8 injected with the Fable 5 prompt diverges across the board in brand tone, copy voice, and section structure.

Test Results

Jamieson used one and the same instruction — "build a modern Apple-style landing page" — for the comparison.

The result: the same intelligence, yet what the two produced were entirely different species.

• Stock Opus 4.8: a by-the-book generic template
• Opus 4.8 injected with the Fable 5 prompt: carrying Fable 5's carefully tuned aesthetic and expressive rhythm — the layout, the wording, and the way modules are broken down all bear the imprint of its "factory personality"

With nothing but a system prompt swapped, the two pages diverge across the board in brand tone, copy voice, and section structure.

How It Works

The experiment reveals an important finding: a large part of Fable 5's "personality" lives in the system prompt, rather than being determined entirely by the model itself.

A system prompt is essentially a carefully crafted "code of conduct" that defines the model's:

• Output style and tone
• Task decomposition strategy
• Tool-calling logic
• Aesthetic preferences
• Interaction cadence

Using the --system-prompt-file flag, you can inject this code of conduct into any compatible model and get similar behavior.

Safety Notes

The method works, but keep the following in mind:

• --dangerously-skip-permissions is a genuinely high-risk operation: skipping all permission confirmations means the model can freely perform file operations, network requests, and more
• The system prompt file comes from a third-party leak: review its contents yourself before use to make sure it contains no malicious instructions
• This is only a "lite" Fable 5: the gap in the model's underlying capabilities cannot be fully bridged by a prompt
• The method may stop working as Claude Code updates

Background: Why Fable 5 Was Pulled

According to reporting by The Information, the trigger for Fable 5's takedown was an internal test at Amazon (Anthropic's biggest backer):

1. Amazon's team used carefully crafted prompts to bypass Fable 5's safety guardrails
2. CEO Andy Jassy contacted senior officials in Washington directly
3. The White House gave Anthropic an ultimatum: fix the flaw or pull the model
4. Dario Amodei refused to make changes, and the model was forcibly taken down

Fable 5 is essentially Mythos wrapped in a shell of safety guardrails. Once the guardrails failed, the capabilities inside were exposed. It was also the first time the safety-versus-capability paradox of the ASI (artificial superintelligence) era was laid out in the open.

FAQ

• Is this method legal? The leaked system prompt is already public on GitHub, but using --dangerously-skip-permissions is at your own risk.
• Is the effect the same as the real Fable 5? Not exactly — the gap in the model's underlying capabilities cannot be fully closed by a prompt, but output style and task strategy are very close.
• Is there a safer alternative? See OpenRouter's Fusion multi-model blending approach, which uses a Kimi K2.6 + DeepSeek V4 Pro + Gemini 3 Flash combination to achieve results close to Fable 5.