Subagent Creator

Create specialized Claude Code subagents following established best practices.

Shared conventions: See references/creator-conventions.md for the standard creator workflow (gather, generate, quality self-check, validate, review gate, commit).

Prompting best practices: See references/prompt-engineering-guide.md for optimization techniques to apply when generating the agent's system prompt.

Workflow

1. Gather Requirements

Use AskUserQuestion for discrete choices, free-form for open-ended questions.

Agent purpose (free-form):

• What specific tasks will this agent handle?
• What specialized knowledge should it have?
• When should Claude delegate to this agent?

Agent location — AskUserQuestion with header "Location":

• "Project" (.claude/agents/ — versioned with the repo)
• "User Global" (~/.claude/agents/ — available across all projects)
• "Plugin" (agents/ — if adding to a prove-style plugin)

Tool permissions — AskUserQuestion with header "Tools":

• "Read-only (Recommended for reviewers)" (Read, Grep, Glob)
• "Read + Research" (Read, Grep, Glob, WebFetch, WebSearch)
• "Full developer" (Read, Write, Edit, Bash, Glob, Grep)
• "Custom" (let user specify)

Model selection — AskUserQuestion with header "Model":

• "opus" (complex reasoning, architecture decisions)
• "sonnet (Recommended)" (general development, good default)
• "haiku" (simple, repetitive, cost-efficient tasks)

2. Create the Agent File

Agent name — derive a hyphen-case name from the role (e.g., code-reviewer, security-auditor, test-engineer).

Use assets/agent-template.md as the scaffold. Refer to references/subagents-best-practices.md for tool permission patterns and example agents.

Model calibration — adjust the generated system prompt verbosity based on the selected model:

• Opus: Lean instructions, fewer guardrails. Built-in agents are 285-636 tokens — aim for similar.
• Sonnet: Moderate detail, explicit workflow steps.
• Haiku: More scaffolding, explicit examples, precise constraints.

Apply the quality self-check from references/creator-conventions.md before presenting. Key checks for agents:

• Description explains WHAT it does and WHEN to invoke (this drives auto-delegation)
• System prompt body is operational, not pedagogical
• Token budget is appropriate for the model

3. Validate

Before writing, verify:

• Name is hyphen-case and matches filename
• Description explains what it does and when to invoke
• Tools match the agent's role (read-only for reviewers)
• System prompt has clear responsibilities and workflow
• Output format is specified
• Model is appropriate for the task complexity
• If plugin: agents scope exists in .claude/.prove.json

Use AskUserQuestion with header "Review" to confirm: "Create Agent" / "Revise".

Committing

Delegate to the commit skill. Example: feat(agents): add security-auditor agent

Resources

• references/subagents-best-practices.md — tool permissions, example agents, pipeline patterns
• assets/agent-template.md — blank agent template with placeholder fields
• references/creator-conventions.md — shared creator workflow patterns
• references/prompt-engineering-guide.md — prompting techniques for the system prompt