Agent Operating Protocol

A Fable 5-style AGENTS.md and CLAUDE.md template for high-agency coding assistants.

This project is not Claude Fable 5, is not affiliated with Anthropic, and does not reproduce vendor system prompts. It is designed to help coding agents behave more like what developers valued in Fable 5-style agents: proactive investigation, long-horizon planning, tool discipline, verification, fallback behavior, and durable handoff.

Why This Exists

Strong coding agents are not only strong because of model weights. The "Fable 5 feeling" many developers noticed is also an operating protocol: the agent classifies the task, gathers real context, uses tools deliberately, verifies its work, and leaves a recovery path.

This repository turns that style into reusable agent instructions:

• classify the request before acting
• inspect actual files and sources before judging
• use the available tool and capability manifest deliberately
• plan across stages for multi-file or risky work
• verify with tests, type checks, lint, builds, or primary sources
• record durable handoffs instead of relying on chat memory
• protect secrets, user data, and paid or external tools by default
• recover from missing tools instead of stopping at a dead end

This repository gives you a clean starting point for that protocol.

What "Fable 5-Style" Means Here

"Fable 5-style" means high-agency coding behavior inspired by public discussion of Claude Fable 5 and similar frontier coding agents. It does not mean:

• an official Anthropic prompt
• a leaked or reconstructed system prompt
• a Fable 5 clone
• a model replacement
• a jailbreak or policy bypass

In this repo, the useful part is the operating pattern: routing, reconnaissance, planning, verification, fallback, and durable handoff.

Included Files

• AGENTS.md - shared protocol for Codex, coding agents, and repo-local agent instructions.
• CLAUDE.md - Claude Code bootstrap that imports or mirrors AGENTS.md and adds Claude-specific guidance.
• examples/AGENTS.project.md - a project-local overlay example.
• examples/CLAUDE.project.md - a Claude Code project overlay example.
• scripts/install.sh - optional dry-run-first installer for global Codex and Claude Code locations.
• NOTICE.md - attribution and non-affiliation notes.

Quick Start

For a repo-local setup, copy the templates into your project root:

cp AGENTS.md /path/to/your/repo/AGENTS.md
cp CLAUDE.md /path/to/your/repo/CLAUDE.md

For global use, inspect the files first, then run the installer:

./scripts/install.sh
./scripts/install.sh --apply

The installer defaults to dry-run mode. With --apply, it backs up existing files before writing:

• ~/.codex/AGENTS.md
• ~/.claude/CLAUDE.md

Positioning

This is a public, vendor-neutral operating protocol. It is inspired by public discussions of high-agency coding agents and public model announcements, including Anthropic's public Fable/Mythos announcements, but it does not include or claim to reconstruct any private system prompt.

Do not present this project as:

• an official Anthropic, OpenAI, or vendor prompt
• a leaked system prompt
• a Fable 5 replacement model
• a jailbreak or policy bypass

It is a practical engineering template.

Useful public context:

• Anthropic's Claude Fable 5 / Mythos 5 announcement: https://www.anthropic.com/news/claude-fable-5-mythos-5
• Anthropic's access suspension statement: https://www.anthropic.com/news/fable-mythos-access

Customization

Keep global files focused on durable behavior:

• language and communication style
• planning and verification rules
• git and destructive-command boundaries
• external API and data exposure rules
• fallback and handoff expectations

Keep project-local files focused on project facts:

• architecture
• commands
• test suite
• deployment rules
• module boundaries
• release process

License

MIT. See LICENSE.