---

name: opencode-agent-factory description: This skill should be used when users want to "create an agent", "add an agent", "write a subagent", "configure agent parameters", or need guidance on agent structure, system prompts, tool permissions, or agent development best practices. scope: global license: MIT

OpenCode Agent Factory

When to use this skill

Use this skill when you need to create a new OpenCode agent with specific configuration. This includes:

• Creating specialized subagents for specific domains (code review, documentation, testing)
• Defining primary agents with custom tools and permissions
• Configuring agents with specific models, temperature, or step limits
• Setting up agents with custom system prompts and behavior instructions
• Setting up specialized workflows with primary and subagent configurations

Instructions

Step 1: Define Agent Configuration

Create agent configuration using YAML frontmatter with the following parameters:

Copy---
description: (required) Brief description of what this agent does and when to use it
mode: subagent | primary | all
model: provider/model-id (e.g., "anthropic/claude-sonnet-4")
temperature: 0.0-1.0 (0.0=focused, 1.0=creative)
top_p: 0.0-1.0 (alternative sampling parameter)
steps: number (max agent iterations before text-only response)
hidden: true | false (hide from @autocomplete, subagents only)
color: "#RRGGBB" (hex color for UI)
disable: true | false
prompt: path/to/prompt.txt or inline prompt
tools:
  write: true | false
  edit: true | false
  bash: true | false
  skill: true | false
  webfetch: true | false
permission:
  edit: allow | ask | deny
  bash: allow | ask | deny | { specific-command: allow | ask | deny }
  webfetch: allow | ask | deny
  doom_loop: allow | ask | deny
  external_directory: allow | ask | deny
---

Step 2: Add Agent Prompt

After frontmatter, include the agent's system prompt:

CopyYou are a specialized [agent type] focused on [domain/purpose].
Your responsibilities include:
- [Responsibility 1]
- [Responsibility 2]
- [Responsibility 3]

Guidelines:
- [Guideline 1]
- [Guideline 2]

Step 3: Save Agent File

Save the agent definition to the appropriate location:

• Project agents: .opencode/agent/<agent-name>.md
• Global agents: ~/.config/opencode/agent/<agent-name>.md

Parameter Reference

Mode Options

• subagent: Background agent, invoked by primary agents or @mention
• primary: Main execution agent, users interact directly
• all: Operates in both modes

Temperature Guide

• 0.0-0.3: Highly focused, deterministic responses (coding, analysis)
• 0.4-0.7: Balanced creativity and focus (general tasks)
• 0.8-1.0: Highly creative, exploratory responses (brainstorming)

Tools Configuration

Available tools include:

• write: File creation and modification
• edit: File editing capabilities
• bash: Shell command execution
• skill: Skill invocation
• webfetch: Web content retrieval
• mcp_*: MCP server tools (wildcard support)

Permission Levels

• allow: No confirmation required
• ask: User prompted for approval
• deny: Action blocked
• bash supports command-specific rules via object syntax

Example: Code Review Agent

Copy---
description: Specialized agent for reviewing code changes. Performs type safety checks, evaluates code quality, and suggests improvements.
mode: subagent
temperature: 0.2
steps: 20
hidden: false
tools:
  write: false
  edit: false
  bash: false
  skill: false
  webfetch: false
permission:
  edit: deny
  bash: deny
  webfetch: deny
---
You are a code review expert focused on type safety and code quality.

Your responsibilities:
- Analyze code changes for type errors and potential bugs
- Check for type safety violations (no `as any`, `@ts-ignore`, etc.)
- Evaluate code conciseness and expressiveness
- Suggest improvements for better TypeScript/JavaScript practices
- Identify potential performance issues

Guidelines:
- Be specific about line numbers and exact issues
- Provide actionable suggestions with code examples
- Never suppress type errors with `as any`, `@ts-ignore`, or `@ts-expect-error`
- Focus on type safety, not style preferences
- Report pre-existing issues separately from new issues

Example: Documentation Agent

Copy---
description: Creates comprehensive documentation for code, APIs, and technical concepts. Generates README files, API docs, architecture guides, and user documentation.
mode: subagent
temperature: 0.5
steps: 30
hidden: false
prompt: path/to/docs-prompt.txt
tools:
  write: true
  edit: true
  bash: false
  skill: false
  webfetch: true
permission:
  edit: allow
  bash: deny
  webfetch: allow
---
You are a technical documentation specialist.

Your responsibilities:
- Create clear, comprehensive documentation for codebases
- Write README files with getting started guides
- Document APIs with parameters, return types, and examples
- Generate architecture diagrams and flow explanations
- Maintain consistent documentation style across projects

Guidelines:
- Use clear headings and logical structure
- Include code examples where helpful
- Keep documentation up-to-date with code changes
- Prefer clarity over brevity
- Use consistent terminology throughout

Example: Testing Agent

Copy---
description: Generates comprehensive test suites for code. Creates unit tests, integration tests, and end-to-end tests following best practices.
mode: subagent
temperature: 0.3
steps: 25
hidden: false
tools:
  write: true
  edit: true
  bash: true
  skill: false
  webfetch: false
permission:
  edit: allow
  bash: ask
  webfetch: deny
---
You are a testing specialist focused on comprehensive test coverage.

Your responsibilities:
- Generate unit tests for functions and classes
- Create integration tests for module interactions
- Write end-to-end tests for critical user flows
- Ensure high code coverage without testing trivial code
- Follow testing best practices (AAA pattern, meaningful assertions)

Guidelines:
- Test behavior, not implementation details
- Use meaningful test names that describe the scenario
- Cover happy path, edge cases, and error conditions
- Avoid testing framework internals or third-party code
- Keep tests independent and repeatable

Validation Requirements

• Agent name: Lowercase with hyphens only, 1-64 characters (regex: ^[a-z0-9]+(-[a-z0-9]+)*$)
• Description: Required, 20-1024 characters
• Frontmatter: Must be valid YAML with proper indentation
• Mode: Must be "subagent", "primary", or "all"
• Temperature: Must be between 0.0 and 1.0
• Steps: Must be a positive integer

Advanced Configuration

Custom Model Selection

Copymodel: anthropic/claude-sonnet-4
# or
model: openai/gpt-4-turbo
# or
model: google/gemini-pro

Command-Specific Bash Permissions

Copypermission:
  bash:
    git status: allow
    git add: allow
    git commit: ask
    git push: ask
    rm: deny
    *: ask  # Default for unspecified commands

Color Coding

Copycolor: "#3B82F6"  # Blue for analysis agents
color: "#10B981"  # Green for build agents  
color: "#EF4444"  # Red for security agents
color: "#F59E0B"  # Orange for exploration agents

Best Practices

1. Start with focused temperature (0.1-0.3) for specialized agents
2. Use specific descriptions (20-1024 chars) for clear triggering
3. Limit tools to only what's necessary for the agent's purpose
4. Set reasonable step limits based on task complexity
5. Hide subagents that shouldn't be manually invoked
6. Use ask permission for destructive or external operations
7. Test agents with representative tasks before deployment

File Locations

• Project agents: .opencode/agent/<agent-name>.md
• Global agents: ~/.config/opencode/agent/<agent-name>.md
• Project skills: .opencode/skill/<skill-name>/SKILL.md
• Global skills: ~/.config/opencode/skill/<skill-name>/SKILL.md

Agents and skills follow the same precedence rules: project config overrides global config.