From tac
Designs custom domain-specific agents with Claude Agent SDK patterns, covering purpose, model selection, prompts, tools, governance, and deployment.
How this skill is triggered — by the user, by Claude, or both
Slash command
/tac:custom-agent-designThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Design domain-specific agents using Claude Agent SDK patterns.
Design domain-specific agents using Claude Agent SDK patterns.
Guide the design of custom agents that solve domain-specific problems with full SDK control over Context, Model, Prompt, and Tools.
Answer these questions:
Output: Purpose statement (2-3 sentences)
Choose based on task complexity:
| Task Type | Model | Why |
|---|---|---|
| Simple transformations | Haiku | Fast, cheap |
| Balanced tasks | Sonnet | Good trade-off |
| Complex reasoning | Opus | Highest quality |
Decision factors:
Choose architecture:
Override (system_prompt=...)
Append (append_system_prompt=...)
System Prompt Template:
# [Agent Name]
## Purpose
[Identity and role definition - 2-3 sentences]
## Instructions
[Core behaviors - bullet list]
- Behavior 1
- Behavior 2
## Constraints
[What the agent must NOT do]
## Examples (if needed)
[Input/Output pairs]
Questions to answer:
Tool configuration:
# Whitelist approach
allowed_tools=["Read", "Write", "Bash"]
# Blacklist approach
disallowed_tools=["WebFetch", "WebSearch", "TodoWrite"]
# No default tools
disallowed_tools=["*"]
# Custom tools
mcp_servers={"domain": custom_mcp_server}
allowed_tools=["mcp__domain__tool1", "mcp__domain__tool2"]
If security/governance required:
hooks = {
"PreToolUse": [
HookMatcher(matcher="Read", hooks=[block_sensitive_files]),
HookMatcher(hooks=[log_all_tool_usage]),
]
}
| Form | Use When |
|---|---|
| Script | One-off automation, ADWs |
| Terminal REPL | Interactive tools |
| Backend API | UI integration |
| Data Stream | Real-time processing |
| Multi-Agent | Complex workflows |
Assemble the ClaudeAgentOptions:
options = ClaudeAgentOptions(
# Context
system_prompt=load_system_prompt("agent_system.md"),
# Model
model="opus",
# Tools
allowed_tools=["Read", "Write", "custom_tool"],
disallowed_tools=["WebFetch", "WebSearch"],
# Custom Tools (if needed)
mcp_servers={"domain": domain_mcp_server},
# Governance (if needed)
hooks=security_hooks,
# Session (if needed)
resume=session_id,
)
When designing a custom agent, provide:
## Custom Agent Design
**Name:** [agent-name]
**Purpose:** [1-2 sentences]
**Domain:** [area of expertise]
### Configuration
**Model:** [haiku/sonnet/opus] - [reason]
**System Prompt Architecture:**
- Type: [Override/Append]
- Reason: [why this choice]
**Tool Access:**
- Allowed: [list]
- Disallowed: [list]
- Custom: [list if any]
**Governance:**
- Hooks: [list if any]
- Security: [considerations]
**Deployment:**
- Form: [script/repl/api/stream/multi-agent]
- Reason: [why this form]
### System Prompt
[Full system prompt content]
### Implementation Notes
[Any special considerations]
| Avoid | Why | Instead |
|---|---|---|
| Competing with Claude Code | Can't beat general agent | Specialize instead |
| Generic system prompts | No domain advantage | Domain-specific |
| Too many tools | Context overhead | Minimal tool set |
| Missing governance | Security risk | Add hooks |
| Override when append works | Loses Claude Code benefits | Use append |
Date: 2025-12-26 Model: claude-opus-4-5-20251101
npx claudepluginhub melodic-software/claude-code-plugins --plugin tacGuides creation of custom agents for Claude Code with specific tool access, behaviors, and task focus. Use when defining agent types or configuring agent capabilities.
Scaffolds custom Claude Agent SDK configurations: generates system prompts, Python agent options, entry points, and docs for new specialized subagents.
Guides creation of custom agents for Claude Code, including file structure, behaviors, tool permissions, workflows, and YAML configuration for specialized tasks.