SkillTool: Execute Skills

When people first see skills like /commit, /verify, and /update-config, they might assume they're just shortcuts for longer prompts.
But looking at the source code, SkillTool is clearly much more complex than a "text alias."

It handles:

• Finding the command associated with the skill
• Parsing arguments
• Handling local skills and MCP skills
• Forking a sub-Agent to run it when necessary

So a more accurate description is:

SkillTool is the skill executor for Claude Code.

tools/SkillTool/SkillTool.ts:

import {
  builtInCommandNames,
  findCommand,
  getCommands,
} from 'src/commands.js'
import { runAgent } from '../AgentTool/runAgent.js'

The file also contains a crucial snippet:

async function executeForkedSkill(
  command: Command & { type: 'prompt' },
  ...
): Promise<ToolResult<Output>> {
  ...
}

This shows that skill execution isn't always completed within the current thread; complex skills can be forked.

User or model triggers skill > SkillTool > finds corresponding command/prompt > parses arguments and metadata > executes in current thread or forks sub-Agent > returns results to main thread

There's a dedicated snippet in the source code:

async function getAllCommands(context: ToolUseContext): Promise<Command[]> {
  const mcpSkills = context
    .getAppState()
    .mcp.commands.filter(
      cmd => cmd.type === 'prompt' && cmd.loadedFrom === 'mcp',
    )
  ...
}

This indicates that SkillTool doesn't just execute local skills; it can also uniformly schedule prompt skills exposed from the MCP side.

1. User inputs /commit
2. SkillTool finds the command for this skill
3. The skill prompt is expanded
4. When necessary, a sub-Agent is forked to run the complete flow
5. Results return to the main thread

SkillTool represents a step toward platformization for Claude Code:

It elevates "skills" from ordinary prompt text into formal capability units that can be scheduled by the system, forked, tracked, and scaled.