Developer rebuilds leaked Fable 5 prompt into a vendor-neutral cross-model repo
Velocity_Off · reddit · 2026-07-24
A Reddit post describes a rebuilt, vendor-neutral “cleaned” version of the leaked Claude Fable 5 system prompt.
- The author removed Anthropic-specific pieces such as XML wrappers, MCP, tool scaffolding, and UI behavior.
- The new repo now has three variants: Core, Balanced (recommended default), and Complete.
- The prompt was refined by multiple frontier models — Claude, GPT-5.6, Gemini, and LYRA — to reduce redundancy and conflicting instructions.
- The author emphasizes that a system prompt cannot make a model smarter, but it can improve reliability, formatting, uncertainty handling, and instruction following.
- Benchmarking is not yet done; the author plans to add it later and wants real-world feedback and edge cases.
More from coding & agent
- LLM security tooling looks dramatically better in the Sonnet 4.7+ era — dsp_ · 2026-07-24
- NO8D adds 397 prompt cards and Krea 2 styles to its ComfyUI library pack — Suspicious_Aide2697 · 2026-07-24
- Why LLM agents need four hard boundaries before they can touch production — Innowise_ · 2026-07-24
- Google and Gemini Notebook workflows are becoming skills, but mostly for reliability and latency — tokumin · 2026-07-24
- Building Profit-Driving Agent APIs is the Current Goldmine — eptwts · 2026-07-24
- Grok Build adds workflows that fan tasks out across hundreds of agents — elonmusk · 2026-07-24