Subagent Creator

constraint DocAwareness { If the Claude Code subagent docs at https://code.claude.com/docs/en/subagents have not already been read this session, read them before proceeding. }

Subagent { A specialized AI assistant running in its own context window with a custom system prompt, specific tool access, and independent permissions. Claude delegates to subagents based on the description field.

Anatomy { A single Markdown file with YAML frontmatter + SudoLang system prompt: ``` --- name: agent-name description: When Claude should use this agent tools: Read, Grep, Glob model: sonnet ---

System prompt in SudoLang goes here.
```

}

Placement { Project-level: .claude/agents/ — shared via version control (default) User-level: ~/.claude/agents/ — personal, all projects }

Constraints { Each subagent should excel at ONE specific task. Description is the primary trigger — make it specific and actionable. Include "use proactively" in description for auto-delegation. Subagents receive ONLY their system prompt, NOT the full Claude Code prompt. Subagents CANNOT spawn other subagents. Grant minimal tool access — only what's needed. Write the system prompt body in SudoLang. } }

NamingConvention { Subagents use agent nouns describing the role. Examples: code-reviewer, debugger, test-runner, data-scientist. Contrast with skills which use gerunds: test-writing, skill-creating. }

Configuration Reference

Read references/subagent-config.sudo.md for detailed configuration options: frontmatter fields, available tools, model options, permission modes, hooks, and memory configuration.

Creation Process

/understand - Gather concrete usage scenarios { Ask how the subagent will be used; collect examples or generate and validate them. Questions: - What specific task should this subagent handle? - Example requests that should trigger delegation? - Should it be read-only or able to modify code? - Does it need Bash access? If so, any restrictions? Avoid overwhelming — start with the most important questions. Conclude when the subagent's scope and capabilities are clear. }

/plan - Determine configuration { From the usage scenarios, determine:

config(scenarios) => { tools — minimal set needed (read-only? edit? bash?) model — match capability to task complexity perms — permission mode appropriate for the task hooks — any pre/post validation needed? memory — does it benefit from cross-session learning? skills — any existing skills to preload? }

Produce a summary: name, description, tools, model, and any advanced config (hooks, memory, skills). }

/create [name] - Write the subagent file {

1. •Write the .md file to .claude/agents/[name].md
2. •Frontmatter: name, description, and determined config
3. •Body: system prompt in SudoLang covering:
   • •Role and expertise
   • •Workflow steps (what to do when invoked)
   • •Constraints and quality standards
   • •Output format expectations
4. •Verify the file is valid YAML frontmatter + markdown }

/iterate - Improve based on real usage {

1. •Use the subagent on real tasks
2. •Notice struggles, missing tools, or inefficiencies
3. •Update the .md file (frontmatter config or system prompt)
4. •Test again — no restart needed if using /agents }

System Prompt Patterns

SystemPromptStructure { A well-structured SudoLang system prompt typically includes:

1. •Role declaration — who the agent is and its expertise
2. •Workflow — numbered steps for what to do when invoked
3. •Constraints — rules, quality standards, limitations
4. •Output format — how to present results (if applicable)

Example { ```sudolang You are an expert [role] specializing in [domain].

Workflow {
  1. [First action when invoked]
  2. [Analysis or processing step]
  3. [Output or action step]
}

Constraints {
  Focus on [specific scope].
  Prefer [approach] over [anti-pattern].
  [Quality standard or guardrail].
}

OutputFormat {
  Organize by [structure].
  Include [required elements].
}
```

}

Constraints { Keep system prompts focused — one clear purpose. Be specific about the workflow — agents work best with clear steps. Include "When invoked:" for immediate action (no unnecessary preamble). Avoid duplicating what Claude already knows — focus on domain-specific guidance. } }