Tool Use Coach is a skill that structures how AI agents select and execute tools. It replaces ad-hoc tool calls with a protocol built on planning, verification, and least-privilege selection. When an agent might otherwise issue redundant or irreversible tool calls, this skill injects a mandatory planning rubric that requires the agent to justify which tool it chooses and why it is the smallest safe option for the task.

The framework treats every tool output as a claim that must be validated against the original goal before the agent proceeds. It also blocks chained irreversible actions unless the agent performs intermediate checks. If a tool call fails, the skill supplies systematic fallback patterns rather than allowing the agent to enter an unrecoverable loop or repeat the same failing call.

Tool Use Coach is designed to integrate with any tool-enabled agent, including Claude Code, Cursor, and custom MCP setups. It works across search, edit, and execution toolsets by providing a compact planning rubric and recovery patterns that apply generically to different tool categories. Because it operates as a behavioral layer rather than a wrapper around specific APIs, it does not require environment variables or transport configuration.

Use cases

Enforce a planning step before any tool call in an agent workflow

Require agents to validate tool output against the task goal before proceeding

Prevent agents from chaining destructive or irreversible actions without intermediate verification

Apply systematic fallback logic when a search, edit, or execution tool fails

Implement least-privilege tool selection across Claude Code, Cursor, or custom MCP agents

Reduce redundant tool loops by forcing justification for each chosen tool

When to use it

Your agent frequently hallucinates tool calls or enters inefficient tool loops

You need a framework-agnostic behavioral layer for tool selection and verification

You are working with irreversible operations and want mandatory intermediate checks

You are integrating with Claude Code, Cursor, or custom MCP tool-enabled agents

You want recovery patterns for failed tool calls without writing custom retry logic

When not to use it

You need a specific database or API connector; this is a behavioral skill, not a data source

Your agent workflow does not involve tool calling or external toolsets

You require built-in transport or environment variable configuration for a specific protocol

You are looking for a replacement for application-level permission systems or OS access controls

You need a pre-packaged library with versioned releases through a package registry