silverdolphin863/claude-efficient-fable

5 stars · Last commit 2026-06-11

Claude Code skill: judge-only orchestration. The frontier model judges, cheap Haiku/Sonnet subagents build and verify, HANDOFF.md remembers.

Code Review Testing

README preview

# claude-efficient-fable

![Efficient Fable](assets/hero.png)

A [Claude Code](https://claude.com/claude-code) skill that turns an expensive frontier-model session (Fable, Opus, or any top-tier model) into a **judge-only orchestrator**: the main session spends tokens exclusively on decomposition, contracts, and verdicts, while cheap Haiku/Sonnet subagents do all the reading, exploring, building, and verifying.

No external tools required. Everything runs on Claude Code's built-in subagents.

## Architecture

![Architecture](assets/architecture.svg)

## The problem

Frontier-model token budgets are tight. Every main-context token spent grepping a codebase, reading a 500-line file, generating boilerplate, or scrolling through test logs is a token not available for the decisions only the frontier model makes well. The final result's quality depends on the judge's decisions and verification gates, not on the judge doing the typing.

## The idea

Three layers:

View full repository on GitHub →