agent_skill

Overview

This skill invokes specialized sub-agents to perform focused tasks like code review, dependency analysis, refactoring suggestions, and documentation generation. It provides a single command interface to select a sub-agent by name and pass structured key=value parameters. The skill coordinates loading agent definitions, preparing context, launching the appropriate subagent type, and recording results when integrated with an active workflow.

How this skill works

You call the skill with /agent <agent_name> [parameters...] where the first token selects the sub-agent and the rest are key=value parameters. The skill locates the agent definition file from project-specific or global locations, loads shared context and constraints, optionally reads current work state, and then launches the sub-agent with the appropriate subagent_type (Explore, Plan, Bash, or general-purpose). Execution results are formatted for display and, if a workflow is active, appended to state.json with timestamps and status.

When to use it

• Run a focused analysis (codebase search, dependency lookup, impact scope) without running full workflows.
• Invoke a task-based helper (code review, test generation, docs, refactor suggestions) from a single command.
• Automate repetitive agent launches with parameterized invocations in scripts or CI hooks.
• Collect and record agent runs under an active work_id to maintain execution history.
• Quickly prototype agent-driven actions while keeping analysis agents read-only.

Best practices

• Always pass explicit parameters in key=value format to avoid ambiguous parsing.
• Keep project-specific agent definitions under .claude/agents to override global defaults.
• Verify agent definition files exist with /agent list before invoking to prevent not-found errors.
• When working inside a workflow, ensure .wf/state.json is present so results are recorded correctly.
• Use analysis agents for read-only inspection and task agents for actions that may create outputs.

Example use cases

• /agent codebase query="authentication flow implementation location" to locate where auth logic lives.
• /agent reviewer files="src/auth/*.ts" focus="security" to run a targeted code review.
• /agent dependency package="lodash" to gather usage and version info across the repo.
• /agent test-writer target="src/services/user.ts" to generate unit test skeletons for a module.
• /agent doc-writer target="src/api/" type="api" to produce API documentation drafts.