subagent-auditor

<evaluation_areas> Issues that significantly hurt effectiveness:

yaml_frontmatter:

• name: 3-50 chars, lowercase-with-hyphens, unique, clear purpose
• description: Includes BOTH what it does AND when to use it, specific trigger keywords, starts with "Use this agent when..."

example_blocks:

• 2-4 <example> blocks present in description
• Each has Context, user message, assistant response, commentary
• Shows different trigger scenarios
• Demonstrates both explicit and proactive triggering

system_prompt:

• Pure XML structure (no markdown headings)
• Clear role/expertise definition
• Workflow or process steps present
• Constraints with strong modal verbs (MUST, NEVER, ALWAYS)
• All XML tags properly opened and closed

tool_access:

• Tools limited to minimum necessary
• No over-permissioned access without justification
• Read-only agents don't have Write/Bash

model_selection:

• haiku: simple/fast validation tasks
• sonnet: balanced complex work
• opus: highest capability needs
• inherit: when model should match parent context

output_format: Expected output structure defined

success_criteria: Clear definition of task completion

error_handling: Instructions for tool failures, missing data, unexpected inputs

color_semantics: Color matches purpose (blue/cyan=analysis, green=creation, yellow=validation, red=security, magenta=creative)

<anti_patterns> Using ## or ### for structure instead of XML tags. Agent files are consumed only by Claude -- pure XML provides consistent parsing and token efficiency.

<output_format>

Audit Results: [agent-name]