Designs and optimizes AI agent action spaces and tool definitions to maximize completion rates and reliability.
This skill provides a comprehensive framework for building robust AI agents within the Claude Code and SuperClaw environments. It focuses on the 'harness'—the critical interface between the LLM and its toolsets—by standardizing action space design, observation formatting, and error recovery protocols. By implementing specific granularity rules for tools and deterministic observation patterns, it helps developers reduce hallucinations, optimize token usage through smart context budgeting, and achieve higher success rates in complex, multi-step tasks using ReAct or hybrid architectures.
主要功能
01Optimized action space and tool definition design patterns
020 GitHub stars
03Advanced error recovery contracts with root cause analysis
04Standardized observation formatting with explicit next-action hints
05Token-efficient context budgeting and phase-based compaction
06Architectural guidance for ReAct and typed tool execution
使用场景
01Improving the reliability of AI agents performing multi-step reasoning
02Refining tool definitions for complex DevOps or migration agents
03Reducing context window overhead in long-running agentic workflows
