mrtooher/fable-mode

What it solves

Fable-mode provides a structured execution discipline for Claude to handle large, complex tasks without skipping critical verification steps or hallucinating success. It prevents the common AI failure mode where a model claims a task "looks right" instead of performing a rigorous, failable check.

How it works

The project implements a "ladder" of skills tuned to different Claude models (Haiku, Sonnet, Opus, and Fable 5.1), each following a core loop:

1. Stage Mapping: Creating a numbered plan with expected outputs and verifiable artifacts for each stage.
2. Named Delegation: Assigning reasoning tasks to Sonnet workers, mechanical tasks to Haiku workers, and verification to read-only verifiers.
3. Failable Verification: Requiring actual tests (e.g., running a command or diffing a file) rather than prose summaries.
4. Self-Critique: Identifying and fixing real weaknesses.
5. Double-Check: Using a panel of fresh agents to attack the finished artifact before delivery.

In version 3, this is structural: an orchestrator agent (with no write access) manages workers and verifiers to ensure delegation actually happens rather than being simulated in prose.

Who it’s for

Users of Claude Code, Cowork, or claude.ai who are performing high-stakes, open-ended research or large-scale mechanical work where accuracy, provenance, and an audit trail are more important than speed.

Highlights

• Model-Specific Tuning: Different skills for different model tiers to guard against their specific known gaps.
• Structural Delegation: Uses dedicated agent definitions to prevent the model from running tasks inline.
• Evidence-Based: Benchmarked against tasks with "unstated traps" to prove it reduces hallucinations in research and increases diligence in mechanical work.
• Execution Guardrails: Includes always-on rules for verify-before-flag and warning batching.