mrtooher/fable-mode

What it solves

It prevents models from rushing into complex, multi-step tasks and shipping errors. It enforces a disciplined, staged execution process that replaces "feeling" with actual verification checks that can fail, ensuring higher quality and more systematic execution on large-scale work.

How it works

The project provides a set of "skills" (prompts/instructions) that force the model to follow a specific procedure: decompose the task into a written stage plan, delegate independent sub-tasks to subagents where possible, perform a verification check at each stage that is designed to fail if the error is present,- and conduct a skeptical self-review before final delivery.

Who it’s for

Users of Claude models (specifically those using environments like Claude Code) who are handling tasks that span multiple files, sources, or sessions and require systematic execution rather than a single-pass attempt.

Highlights

• Staged Execution: Forces decomposition and planning before action.
• Failable Verification: Uses checks that can actually fail rather than relying on the model's intuition.
• Skeptical Self-Review: Includes a mandatory critique phase before delivery.
• Model-Specific Variants: Offers specialized versions for Opus, Sonnet, and Haiku to balance reasoning power and cost.
• Parallel Delegation: Supports routing sub-work to workers when the runtime environment supports agent tooling.