So "leaked" is the viral word, and I kept it because that is what everyone calls it. The accurate version: Anthropic publishes the core prompt itself, and the GitHub file is a community extraction that additionally captures tool instructions the published page only references. Anthropic has not confirmed the extraction is complete or unmodified, and it reflects claude.ai specifically (the API gives you no default system prompt, and Claude Code ships its own instructions).
These come straight from the file, and they work in any CLAUDE.md, project instructions, custom instructions box, or agent prompt.
The prompt is not one wall of rules. Every rule lives inside a scoped section like tone_and_formatting, search_instructions, refusal_handling, or user_wellbeing.
✂️ Steal it: split your instructions into sections like ## voice, ## tools, ## formatting, ## never do. One topic per section, no rule floating homeless at the bottom.
Claude is not told "search when useful". It gets explicit if-then logic: answer directly for stable knowledge, always search for anything that changes (roles, prices, news, new releases), and scale effort to the task (one search for a single fact, several for deep research, suggest the Research feature past that).
⚙️ Steal it: for every tool or recurring choice in your setup, write the trigger and the action. "If the request mentions a brand deal, read the deals folder first. If a stat is older than 6 months, re-verify before using it."
Rules come with demonstrations. There are dedicated example sections, and even a literal "What not to do" list followed by "What this should feel like", with a good phrasing shown next to a weak one.
🧪 Steal it: under each instruction, paste one real example of output you loved and one you rejected, labelled. Models copy examples far more reliably than they follow adjectives like "professional" or "engaging".
Download the extraction as claude-fable-5.md, put it in a folder with your current instructions file, then run this in Claude (Claude Code or a Claude project both work):
Read my current instructions in [your-file.md] and the extracted Claude Fable 5 system prompt in claude-fable-5.md. Rewrite my instructions using Anthropic's structural patterns: 1. Named sections with one job each. 2. Explicit decision rules (trigger plus action) for every tool and recurring choice. 3. At least one worked example per section, including one "what not to do". 4. Severity levels, so hard limits are clearly separated from soft preferences. 5. A short "critical reminders" block at the end repeating only the 3 rules that matter most. Keep all of my original intent. Where my current instructions are ambiguous, list the ambiguity and ask me instead of guessing.
You will see people loading this prompt into open-weight models to make them behave more like Claude. The models named in the viral posts are real and genuinely good: GLM-5.2 by Z.ai (open weights, June 2026, huggingface.co/zai-org/GLM-5.2) and Kimi K2.7-Code by Moonshot AI (open weights, June 2026, huggingface.co/moonshotai/Kimi-K2.7-Code).