Agent Structure

paths:

Investigate existing agents to extract conventions before creating new ones
Frontmatter requires name, description, and tools fields
Use ultrathink. after # Agent Instructions for complex analysis tasks
Include ## Input section when agent needs external context (optional)
Agents return findings; caller applies Pattern Validation and User Selection from CLAUDE.important.md

Section	Purpose
`# Agent Instructions`	Section header (no description needed - use Objectives)
`ultrathink.`	Signal for extended reasoning
`## Objectives`	Bullet point with high-level goal (single item)
`## Process`	Numbered list of workflow steps
`## Important`	Bullet point list of constraints and guidelines
`## Input`	External context (optional)

Use one format per agent, not mixed:

Input Type	Format	Example
Command output	!`command` syntax	!`git diff`
Prompt context	Descriptive text	"The following change summary will be provided:"

Reference

If creating a review agent: Read: files/home/kai/.config/claudex/config/agents/final-review.md

If creating a cleanup agent: Read: files/home/kai/.config/claudex/config/agents/code-cleanup.md

If creating a verification agent: Read: files/home/kai/.config/claudex/config/agents/verification.md