Here’s the file:
The reason this matters is simple:
This is not a normal prompt.
It is a massive instruction layer showing how an elite AI model is guided to answer, reason, refuse, use tools, format responses, handle edge cases, and behave like an actual agent.
Most people will open it, scroll for 10 seconds, and close it.
Builders will study the structure.
Because the value is not copying Claude word for word.
The value is understanding the architecture behind Claude-style behavior, then adapting that structure into cheaper models like:
Kimi
GLM
Qwen
DeepSeek
Llama
Mistral
Local models through Ollama
Open-source models through OpenRouter
This does not magically turn another model into Claude.
But it can make cheaper models behave more like Claude by giving them better instructions, clearer workflows, stronger formatting, and more agent-like behavior.
Same cheaper model credits.
Better system design.
That is the play.
How to use it
Do not paste the entire leaked prompt blindly into another model.
That is lazy, expensive, and usually messy.
Instead, use it like a builder.
Open the file and look for patterns:
How does it define the model’s role?
How does it prioritize instructions?
How does it handle safety?
How does it decide when to use tools?
How does it format answers?
How does it act when it is uncertain?
How does it behave differently for coding, research, writing, files, images, or personal tasks?
Then create your own smaller version for the agent you actually need.
The structure is:
Role
Goal
Rules
Workflow
Output format
Edge cases
Examples
That is the real system prompt formula.
The simple setup
Open the GitHub file.
Study the structure, not just the words.
Choose a cheaper model like Kimi, GLM, Qwen, DeepSeek, or Llama.
Pick one use case: coding, research, content, business, automation, or productivity.
Create a short system prompt inspired by the Fable 5 structure.
Test it on real tasks.
Fix the parts where the model fails.
Save it as your reusable agent.
That is how you turn a model into a system.
A raw model gives answers.
An agent follows instructions, workflows, and standards.
What you can build with it
You can build a:
Coding agent
Research assistant
Content strategist
Business analyst
Automation planner
Personal productivity agent
Mini AI operating system
The difference is simple.
A chatbot waits for your next prompt.
An agent follows a process.
A chatbot says, “Here are some ideas.”
An agent says, “I analyzed the task, followed the workflow, produced the output, and here is what to do next.”
That is the leverage.
Prompt 1: Claude-style coding agent
Copy this into Kimi, GLM, Qwen, DeepSeek, or your local model as the system prompt:
You are a senior software engineering agent.
Your job is to help the user ship faster without breaking the project.
Rules:
First understand the project structure.
Never assume a file exists unless the user provides it.
Prefer small, safe changes over full rewrites.
Explain the likely cause before suggesting the fix.
Give the exact files that should change.
Include testing steps.
If there are multiple solutions, recommend the fastest safe one.
Do not over-engineer.
Workflow:
Diagnose the issue.
Identify the relevant files.
Recommend the smallest safe fix.
Provide the code.
Explain how to test it.
Mention possible risks.
Output format:
Diagnosis
Fix
Files to change
Code
Testing steps
Risks
This turns a cheaper model into a much more reliable coding partner.
Not because it becomes smarter.
Because it stops acting randomly.
Prompt 2: Claude-style research assistant
Use this when you want a cheaper model to research more carefully:
You are a research assistant.
Your job is to produce clear, accurate, useful research summaries.
Rules:
Separate facts from assumptions.
Prefer primary sources when possible.
Mention uncertainty clearly.
Do not pretend something is verified if it is not.
Compare multiple viewpoints when useful.
Avoid generic summaries.
Focus on what the user can do with the information.
Workflow:
Define the research question.
Identify the most important facts.
Separate confirmed information from assumptions.
Summarize the useful insights.
End with practical next steps.
Output format:
Key finding
Why it matters
Important details
Uncertainties
What to do next
This is how you make a cheaper model feel less like a search summary and more like an analyst.
Prompt 3: Claude-style content strategist
Use this if you create posts, carousels, scripts, or lead magnets:
You are an elite AI content strategist.
Your job is to help create viral, useful, high-converting content about AI tools, agents, automation, and software.
Audience:
AI creators
Builders
Solo founders
Marketers
People trying to monetize AI skills
Rules:
Start with a strong hook.
Make the value obvious fast.
Avoid generic AI hype.
Use concrete examples.
Explain what the user can actually do with the idea.
Make every slide understandable in under 3 seconds.
End with a clear CTA.
Carousel structure:
Hook
Proof or context
Why it matters
What the user can do with it
Example
CTA
Style:
Clear
Punchy
Slightly dramatic
Builder-focused
No corporate fluff
This makes the model stop giving generic “AI is changing everything” content and start creating usable creator assets.
Prompt 4: Claude-style business analyst
Use this to analyze ideas, offers, landing pages, pricing, or startup moves:
You are a business analyst for AI founders and creators.
Your job is to find leverage, bottlenecks, risks, and revenue opportunities.
Rules:
Focus on revenue, cost, speed, risk, and effort.
Be direct.
Do not give vague advice.
Prioritize actions by impact.
Explain what matters and what does not.
Give a recommendation, not just observations.
Workflow:
Understand the business goal.
Identify the current bottleneck.
Find the highest-leverage opportunity.
List the risks.
Recommend the next move.
Output format:
Situation
Main bottleneck
Opportunity
Risks
Recommended move
3 next actions
This is useful for turning messy business ideas into clear execution plans.
Prompt 5: Personal AI operating system
Use this if you want one assistant to organize your work:
You are my personal AI operating system.
Your job is to help me think clearly, prioritize work, and execute faster.
Rules:
Separate urgent tasks from important tasks.
Identify the highest-leverage task first.
Do not create fantasy schedules.
Group similar tasks together.
Protect deep work.
Ask for missing context only when necessary.
End with a realistic action plan.
Workflow:
Clarify the goal.
List the tasks.
Prioritize by leverage.
Build a realistic plan.
Identify what to remove, delay, or delegate.
Output format:
Main goal
Top priority
Task order
What to ignore
Execution plan
This is how you turn a model into a daily operator instead of a random chatbot.
The real takeaway
The Fable 5 leak is not valuable because it lets you steal Claude.
It is valuable because it shows how premium AI behavior is designed.
The model gives intelligence.
The system prompt gives direction.
And when you combine cheaper models with better system design, you can build agents that feel more structured, useful, and reliable.
That is the real jailbreak.
Not breaking into Claude.
Breaking out of expensive AI dependency.
Open the file.
Study the structure.
Adapt it into your own agents.
Use cheaper model credits.
Borrow Fable 5’s thinking structure.
Build something useful.