claude-fable-5-system-prompt-clean: A Vendor-Neutral Rewrite of the Claude Fable 5 System Prompt

The Problem Is Prompt Portability, Not Prompt Quality

Most teams that lift a system prompt from one vendor and paste it into another discover the same thing: a large fraction of the text describes infrastructure the new model does not have. The README states the original Claude Fable 5 / Mythos 5 prompt contains XML conventions, tool interfaces, product-specific behavior and internal implementation guidance that exist for Claude's production environment. Reused elsewhere, those segments consume context and give the model instructions it cannot act on. This repository answers a narrower question, quoted from its own philosophy section: which behavioral principles remain valuable after removing implementation details specific to Claude itself. The audience is anyone maintaining one behavioral prompt across several frontier models, plus people writing custom instructions, Gems or custom GPT system prompts where a large pasted block is unwelcome. It is not aimed at teams building Claude-native tooling, since the material they need is exactly what was removed.

Three Files, One Trade-Off Axis

The repository ships three prompt files rather than one, and the difference between them is size against coverage. universal/core.md holds core behavioral principles with minimal token usage. universal/balanced.md is described as recommended for most use cases, balancing prompt size with behavioral coverage. universal/complete.md retains nearly all vendor-neutral behavioral guidance while staying model-agnostic. That is the whole architecture: there is no build step, no runtime, no loader and no configuration schema. The repository is text plus a README in English, Simplified Chinese and Japanese. The consequence is that adoption cost is close to zero and so is the switching cost, which also means there is nothing here to version-pin in a dependency file. If you need reproducibility, you pin a commit hash and copy the file into your own repository, because the upstream file can change under you.

What the Filter Actually Removes, and What Survives

The removal list is explicit. Excluded: XML wrappers and formatting conventions, MCP and tool-specific infrastructure, product-specific instructions, internal UI behavior, storage and artifact management, and environment-specific implementation details. The preserved list is equally explicit: understanding user intent before responding, honest handling of uncertainty, clear communication and natural writing, high-quality code generation, practical recommendations with trade-off analysis, verification before finalizing responses, objective error correction, and avoiding unnecessary meta-commentary. Read those two lists together and the design intent is clear. Everything that describes a capability was cut; everything that describes a disposition was kept. That is a defensible split, but note what it implies: a prompt made only of dispositions cannot tell a model which tools exist, so any agent that needs to call functions, read files or emit structured output must supply that layer itself. This repository will not do it for you.

Getting It Running: Paste, Do Not Install

There is no package to install. The README gives two usage paths. For API platforms, paste the contents of the desired prompt into the model's system prompt or system instructions field. For consumer interfaces, the same text can be used as custom instructions, a custom GPT system prompt, a Gem, or an equivalent feature where the product supports one. In practice that means reading universal/balanced.md, copying it, and pasting it into the system field of your provider's API call or console. There are no config keys, no environment variables and no CLI entry point documented, so any automation is whatever you build around the file. If you are running several models, the practical pattern is to keep the Markdown in version control next to your application code and read it at request time, which at least makes prompt changes reviewable in a pull request instead of living in a web form.

What You Give Up When You Strip the Infrastructure

The largest limitation is stated by the project itself and then left unresolved: this is a filter, not a port. Removing MCP and tool infrastructure, artifact and storage handling, and product-specific instructions means the output is not a drop-in replacement for the original in any environment that used those features. A Claude deployment relying on XML conventions for structured output will find that guidance gone. A second limitation is that the repository publishes no evaluation. The README claims the retained guidance generalizes well across modern models, but there is no comparison, no scoring and no example of a prompt before and after. The compatibility list (Claude, ChatGPT, Gemini, DeepSeek, Qwen, Grok, GLM and other instruction-following LLMs) is a statement of intent, not a result. Third, the three variants are described qualitatively, with no token counts given, so choosing between core, balanced and complete means opening the files and measuring them yourself. Finally, the project does not claim to improve on Anthropic's original, and readers should take that at face value: if the original works in your stack, replacing it with a reduced version is a downgrade unless portability is the thing you need.

How This Differs From Keeping the Original Prompt

The obvious alternative is not a competing repository, it is the original Anthropic prompt left intact. The difference is approach rather than feature set. Keeping the original preserves everything: the XML conventions, the tool interfaces, the product behavior, the internal guidance. It also means carrying instructions that a Gemini or Qwen endpoint cannot execute, and paying context for them on every request. This repository takes the opposite position, that a prompt should describe how a model reasons and writes rather than what a particular vendor's runtime exposes, and it accepts the loss of platform features as the price. Neither position is wrong; they answer different questions. If your work is Claude-only and uses tools, artifacts or MCP, the original is the correct artifact and this repository is the wrong one. If your work spans several providers and only needs behavioral consistency, the filtered version is the smaller thing to maintain.

Maintenance Cost, Licence and What to Check First

The repository is MIT licensed, which permits reuse, modification and redistribution provided the licence and copyright notice are retained. That is a permissive arrangement, but it covers the repository's own text only. The underlying Claude Fable 5 / Mythos 5 system prompt originates with Anthropic, and the README does not state under what terms that source material was published or what attribution the adaptation carries. Anyone shipping these prompts inside a commercial product should resolve that question with their own counsel rather than treating the MIT badge on the repository as covering the upstream text. On maintenance, the cost is low and mostly manual: no releases have been published, so there is no changelog to follow, and updates arrive as commits to the default branch. The README invites issues and pull requests for vendor-specific behavior that should be generalized, redundant instructions that can be simplified, or compatibility improvements. Practically, that means you should vendor the file into your own repository at a pinned commit, note which of the three variants you chose, and re-diff when you decide to move forward.

Editorial conclusion

Adopt it if you are moving a behavioral prompt across Gemini, ChatGPT, DeepSeek or Qwen and want the Claude-specific scaffolding gone before you paste anything. Skip it if you depend on tool interfaces, XML conventions, artifacts or MCP wiring, because all of that was deliberately deleted. Before adopting, open universal/balanced.md and universal/complete.md side by side and count what complete.md adds; then diff the retained section against the original Anthropic prompt you are migrating from, since the repository does not publish that diff itself.

Official sources

1. Issues
2. KinetiNode/claude-fable-5-system-prompt-clean on GitHub
3. License: MIT
4. README