agent-improver

name	agent-improver
description	Create and improve agents. Trigger when user says: "create an agent for X", "build me an agent", "I need an agent that does Y", "make a code review agent", "create a testing agent", "improve my agent", "list agents", "show me agents", "what agents do I have".

Agent Management

Create Agent

Write directly to .claude/agents/{name}.md:

---
name: agent-name
description: >-
  What this agent does and when Claude should delegate to it.
  Be specific. Include "Use proactively" if it should auto-trigger.
tools: Read, Grep, Glob, Bash
# Optional fields below - include only when needed:
# disallowedTools: Write, Edit           # Tools to deny (removed from inherited set)
# model: sonnet                           # sonnet, opus, haiku, inherit, or full model ID
# permissionMode: default                 # default, acceptEdits, dontAsk, bypassPermissions, plan
# maxTurns: 20                            # Max agentic turns before stopping
# skills:                                 # Skills preloaded into context at startup
#   - api-conventions
#   - error-handling
# mcpServers:                             # MCP servers scoped to this agent
#   - github                              # Reference existing server by name
#   - playwright:                         # Or inline definition
#       type: stdio
#       command: npx
#       args: ["-y", "@playwright/mcp@latest"]
# hooks:                                  # Lifecycle hooks scoped to this agent
#   PreToolUse:
#     - matcher: "Bash"
#       hooks:
#         - type: command
#           command: "./scripts/validate.sh"
# memory: project                         # Persistent memory: user, project, or local
# background: false                       # true = always run as background task
# effort: high                            # low, medium, high, max
# isolation: worktree                     # Run in isolated git worktree
---

You are [role description]. [System prompt for the agent.]

When invoked:
1. First step
2. Second step
3. Third step

## Output Format

[Expected output structure]

Frontmatter Reference

Field	Required	Purpose
name	Yes	Unique ID, lowercase + hyphens
description	Yes	When Claude should delegate. Include "Use proactively" for auto-trigger
tools	No	Allowlist of tools. Inherits all if omitted
disallowedTools	No	Denylist (removed from inherited/specified set)
model	No	sonnet, opus, haiku, inherit, or full model ID (e.g. claude-opus-4-6)
permissionMode	No	default, acceptEdits, dontAsk, bypassPermissions, plan
maxTurns	No	Max agentic turns before auto-stop
skills	No	Skills preloaded into agent context (full content injected)
mcpServers	No	MCP servers: string reference or inline definition
hooks	No	PreToolUse/PostToolUse/Stop hooks scoped to this agent
memory	No	user (~/.claude/agent-memory/), project (.claude/agent-memory/), local (.claude/agent-memory-local/)
background	No	true = always run as background task
effort	No	low, medium, high, max
isolation	No	worktree = isolated git worktree copy

Agent Scope

Location	Scope	Priority
--agents CLI flag	Current session only	1 (highest)
.claude/agents/	Current project	2
~/.claude/agents/	All your projects	3
Plugin agents/ dir	Where plugin enabled	4 (lowest)

Higher priority wins when names collide.

Model Selection Guide

Model	Best for
haiku	Fast, cheap: exploration, search, simple tasks
sonnet	Balanced: code review, analysis, moderate complexity
opus	Most capable: complex reasoning, architecture, planning
inherit	Same as main conversation (default)

Memory Scopes

Scope	Location	Use when
user	~/.claude/agent-memory/{name}/	Knowledge applies across all projects
project	.claude/agent-memory/{name}/	Project-specific, shareable via VCS
local	.claude/agent-memory-local/{name}/	Project-specific, NOT in VCS

When memory is enabled, include in system prompt:

Update your agent memory as you discover patterns, conventions, and
architectural decisions. This builds institutional knowledge across conversations.

Key Principles

1. Focused responsibility — Each agent excels at one specific task
2. Detailed description — Claude uses this to decide when to delegate
3. Minimal tool access — Grant only necessary permissions
4. Imperative system prompt — Clear instructions, not persona descriptions
5. Specify model — Use haiku for fast/cheap, sonnet for balanced, opus for complex

List Agents

ls .claude/agents/*.md

Add Agent to Chain Flow

Edit the active chain file in .claude/chains/, add to flow: section:

flow:
  new-agent:
    routes:
      next-agent: "Description of when to go here"
      fallback-agent: "Description of fallback"
    decide: |
      If SUCCESS → "next-agent"
      If FAILED → "fallback-agent"

Then update the previous agent's routes to point to the new agent.

Agent Teams vs Parallel Subagents

IMPORTANT: When user wants agents to work in parallel, choose the right pattern:

Agent Team (TeamCreate) — for complex parallel work

• Multiple independent Claude Code sessions
• Teammates have their own context, can message each other
• Shared task list with self-coordination
• Use when: teammates need to discuss, challenge each other, or coordinate
• Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

To create an agent team, use TeamCreate tool (NOT the Agent tool):

TeamCreate(team_name: "review-team", description: "Parallel code review")

Then spawn teammates via Agent tool with team_name parameter:

Agent(subagent_type: "code-reviewer", team_name: "review-team", name: "quality-reviewer")
Agent(subagent_type: "security-scan", team_name: "review-team", name: "security-reviewer")

When done, shut down teammates via SendMessage, then clean up:

TeamDelete()

Parallel Subagents (Agent tool) — for quick independent tasks

• Run within one session, report results back to parent
• Cannot message each other
• Use when: tasks are independent and only the result matters

Agent(subagent_type: "code-reviewer", run_in_background: true)
Agent(subagent_type: "security-scan", run_in_background: true)

Chain Team Node — for chain workflows

In the active chain file (.claude/chains/), use type: team to automatically trigger TeamCreate:

flow:
  parallel-review:
    type: team
    teammates:
      - name: code-quality
        agent: code-reviewer
        prompt: "Review code for bugs and quality"
      - name: security-check
        agent: security-scan
        prompt: "Scan for security vulnerabilities"
    routes:
      next-agent: "All reviews passed"
      fallback-agent: "Issues found"
    decide: |
      If ALL passed → "next-agent"
      If ANY failed → "fallback-agent"

The chain hook handles TeamCreate/TeamDelete automatically.

Decision Guide

User says	Use
"run review and security in parallel"	Agent Team (TeamCreate)
"create a team of agents"	Agent Team (TeamCreate)
"agents that collaborate/discuss"	Agent Team (TeamCreate)
"quick parallel search"	Parallel Subagents (Agent tool)
"add parallel step to chain"	Chain team node (type: team)

DO NOT create multiple regular agents via Agent tool when user asks for a "team" — use TeamCreate instead.

Send Messages to Agent

Use Memory Bus:

mcp__brainbrew__memory_add(content: "...", target: "agent:agent-name")

Improve Existing Agent

1. Read the agent file: .claude/agents/{name}.md
2. Check against frontmatter reference above — are fields correct?
3. Verify description is specific with trigger phrases
4. Verify tools are scoped to minimum needed
5. Verify system prompt uses imperative voice
6. Check if skills field should preload companion skills
7. Check if memory would benefit this agent