SHART (Save Half AI Request Tokens) is a lightweight, repo-local agent framework for Claude Code.
It preserves the reliability of GSD-style agentic workflows—planning, execution, verification, and safety—while dramatically reducing token usage through file-based memory, strict verbosity limits, and deterministic execution rules.
SHART is designed for developers who want:
- Autonomous multi-step execution
- Minimal narration and reflection
- Deterministic, verifiable changes
- Predictable token consumption
- Repo-enforced agent behavior (not prompt sprawl)
This repository contains:
- A minimal system prompt for Claude Code
- Token, safety, and execution policies
- Cheap vs expensive execution modes
- Optional planner/executor split
- CI enforcement to prevent workflow drift