This repository distills the Fable Workflow, a structured problem-solving methodology from Claude Fable 5, into executable skills for AI agents. It addresses common agent failures like poor classification, undefined success, and dishonest reporting. Targeted at engineers and researchers, it enhances agent reliability and trustworthiness via a rigorous "think / act / prove" framework with adversarial evaluation.

How It Works

The Fable Workflow follows a "think / act / prove" philosophy: think (fable-method), act (fable-loop), prove (fable-judge). Core skills include classifying the ask, defining success via named verification, gathering parallel primary evidence, committing to one recommendation, performing surgical edits, verifying by observation, and reporting outcomes first with honest caveats. Rules were iteratively refined through 159 adversarial agent runs, each justified by a specific failure mode. Domain adapters extend the loop beyond coding to sectors like marketing, research, and finance, defining evidence requirements and verification methods.

Quick Start & Requirements

• Installation: Recommended via Claude Code plugin (/plugin marketplace add Sahir619/fable-method). Standalone installation uses git clone and install.sh/install.ps1. Proactive use configurable in ~/.claude/CLAUDE.md.
• Prerequisites: Requires an agent/model capable of executing skills. No specific hardware/software versions mandated beyond the agent environment.
• Links: Plugin install: /plugin marketplace add Sahir619/fable-method. Repo: https://github.com/Sahir619/fable-method.

Highlighted Details

• Adversarial Verification: fable-judge rigorously verifies work by diffing changes and executing checks, combating reward hacking and false completion claims.
• Rule-Based Design: Every rule addresses observed failures, ensuring robustness against specific agent weaknesses (e.g., surfacing spec-vs-test conflicts).
• Domain Adaptability: Seven domain adapters tailor the workflow for marketing, research, business, finance, legal, and design, enforcing sector-specific evidence and verification.
• Performance: Ties or out-ranks bare frontier models on mixed tasks; demonstrably flags planted frauds and spec-vs-test conflicts.

Maintenance & Community

A community distillation, not an official Anthropic artifact, originating from work with Claude Fable 5. No specific community channels or maintainer details are provided.

Licensing & Compatibility

• License: MIT.
• Compatibility: Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The method structures agentic work but does not enhance a model's raw knowledge. Its value concentrates on mitigating specific failure modes ("traps") rather than improving simple, attended tasks. Medical/clinical work is excluded. Evaluation is "smoke-test grade."