Skill for improving tool use accuracy. Monitors tool selection decisions, detects errors (wrong tool, missing params, type mismatches, ambiguity), and applies learned prevention rules to improve future accuracy.
This skill enables closed-loop learning for tool use accuracy. It monitors agent tool selection decisions and learns from mistakes.
User Request → Agent picks tool → Execution → Success/Failure
↓ ↓
Prevention Rules ← Learning Engine ← Error Record
↓
(warnings added to context)
↓
Agent sees warnings → picks better tool
This skill works with CannyForge adapters (LangGraph, MCP) that wrap agents. The skill itself is declarative — it defines what errors to track and what context fields to use. The actual tool selection is done by the agent.
Example integration (LangGraph):
from cannyforge.adapters.langgraph import CannyForgeMiddleware
middleware = CannyForgeMiddleware(forge, skill_name="tool_use")
agent = create_agent(middleware=[middleware], tools=[...])