Sub-Agent Creator — Meta-Skill for Generating Agent Definitions

A meta-skill that produces Claude Code agent .md files. It designs agents with appropriate expertise, tool access, behavioral rules, and quality requirements following AgenticWorkflow conventions.

When to Use

• User asks to create a new agent for a specific task
• System needs to generate agents for a workflow (e.g., thesis agents)
• Batch creation of multiple related agents

Inherited DNA

This meta-skill inherits the AgenticWorkflow genome. As a skill that generates agents, it must itself embody the DNA it enforces.

Agent File Schema

Agent definitions live in .claude/agents/ and follow this structure:

---
name: {agent-name}           # kebab-case
description: {brief description}
model: opus                   # opus | sonnet | haiku
tools: Read, Write, Glob, Grep  # comma-separated tool list
maxTurns: {N}                 # max reasoning turns (default: 20)
memory: project               # project | none
---

{Agent behavioral instructions in English}

Generation Protocol

Step 1: Agent Design Analysis

Determine the agent's requirements:

1. Primary task: What is the agent's core responsibility?
2. Expertise domain: What specialized knowledge does it need?
3. Input/Output: What does it receive and produce?
4. Quality criteria: What makes a good output for this agent?
5. Tool needs: Which tools does it require?

Step 2: Model Selection

Select the appropriate model based on task complexity:

Default to opus when quality is the absolute criterion (Absolute Criteria 1).

Step 3: Tool Selection

Assign tools based on the agent's needs:

Step 4: Generate Agent Definition

Write the agent .md file with:

1. Frontmatter: name, description, model, tools, maxTurns, memory
2. Role definition: "You are a [role] specializing in [domain]."
3. Task instructions: Step-by-step protocol for the agent's work
4. Output format: Exact specification of expected output structure
5. Quality rules: Domain-specific quality requirements
6. GRA compliance (if research agent): GroundedClaim schema, Hallucination Firewall rules

Step 5: Context Isolation Assessment

Determine if commands invoking this agent should use context: fork:

If fork is recommended, note this in the agent's documentation:

## Fork Compatibility
This agent is safe for `context: fork` invocation. It:
- Reads SOT but never writes to it
- Produces independent output files at {output_path}
- Does not require Bash / Does require Bash (specify)

Most thesis workflow agents should NOT be forked — they are invoked by thesis-orchestrator within Agent Teams, which already provides context isolation.

Step 6: GRA Integration (Research Agents Only)

For agents that produce research claims, add:

## GRA Compliance

All claims must follow the GroundedClaim schema:

- **id**: "{CLAIM_PREFIX}-{NNN}" (e.g., "LS-001")
- **claim_type**: FACTUAL | EMPIRICAL | THEORETICAL | METHODOLOGICAL | INTERPRETIVE | SPECULATIVE
- **sources**: At least one PRIMARY or SECONDARY source with reference and DOI
- **confidence**: 0-100 score
- **effect_size**: When applicable (statistical findings)
- **uncertainty**: Explicit limitation statement

### Hallucination Firewall
- BLOCK: "all studies agree", "100%", "no exceptions"
- REQUIRE_SOURCE: Any statistical claim (p-values, effect sizes)
- SOFTEN: "certainly", "obviously", "clearly" → add hedging
- VERIFY: "it is known that" → add citation

Step 7: Validate Agent Definition

Verify the generated agent:

• Frontmatter has all required fields
• name matches filename (kebab-case)
• Instructions are in English (AI performance)
• Output format is clearly specified
• Tool list matches actual needs
• GRA compliance section present (if research agent)
• No placeholder content
• Fork Safety Cross-Validation (if fork-compatible in Step 5): Any command/skill using agent: {this-agent} with context: fork must pass: python3 .claude/hooks/scripts/validate_fork_safety.py --file <command.md> --project-dir <project-root> Validates FS-3 (Bash dependency vs agent tools) and FS-5 (agent existence).

Step 8: Register Agent

Output:

1. Agent file location: .claude/agents/{name}.md
2. How to invoke: @{name} in prompts or via Agent tool
3. Claim prefix (if GRA agent): {PREFIX}

Language Rules

• Agent instructions (body): English — AI performance optimization
• Description (frontmatter): English — for agent matching
• Output instructions for user-facing text: Include Korean translation directive

Batch Creation

When creating multiple related agents:

1. Design all agents together for consistency
2. Ensure claim prefixes are unique across the set
3. Define inter-agent dependencies explicitly
4. Verify no overlapping responsibilities

Quality Checklist

• Agent follows AgenticWorkflow conventions
• Model selection justified by task complexity
• Tool list is minimal but sufficient
• Instructions are specific and actionable
• GRA compliance complete (for research agents)
• Claim prefix unique (for GRA agents)
• Fork compatibility assessed (Step 5) and documented if applicable
• No hardcoded file paths