Fable 5 cleaned prompt v2

As some of you may remember from my previous post, I released a shortened version of the leaked Claude Fable 5 system prompt by removing Anthropic-specific infrastructure (XML, MCP, tool wrappers, UI behavior, etc.) that had little or no value on other models. After reading a lot of your feedback, I agreed that the first version wasn't where I wanted it to be. So I rebuilt it from the ground up. This time I used multiple frontier models (Claude, GPT-5.6, Gemini, and LYRA) to critique the prompt, identify redundancy, find conflicting instructions, and improve its cross-model behavior. The repository now contains three variants: Core — Minimal token overhead while preserving the highest-impact behavioural guidance. Balanced — My recommended default, includes most vendor-neutral behavioural guidance without unnecessary bloat. Complete — The most comprehensive version, covering reasoning, writing, coding, reliability, document fidelity, instruction precedence, and more. Before anyone says "a prompt can't make a model smarter", I know. A system prompt cannot increase a model's intelligence, unlock hidden capabilities, or magically improve benchmarks. What it can do is influence how the model uses the capabilities it already has. A well-designed prompt can help reduce hallucinations, improve instruction following, encourage better uncertainty handling, produce more consistent formatting, generate more complete code, and generally make responses more predictable and reliable. The goal of this project isn't to "upgrade" GPT, Claude, Gemini, or any other model and magically turn it into Fable 5.The goal is to extract the vendor-neutral behavioral principles from a very large, model-specific system prompt and package them into lightweight, portable prompts that work well across modern LLMs. As always, feedback is welcome—especially benchmark results, edge cases, and examples where a prompt underperforms. Empirical testing is far more valuable than subjective opinions, and I'd love to keep improving the project based on real-world results. as for official benchmarks.. im working on other projects right now and don't have time to create the benchmarks but i will add that to the repo eventually.