plugin-dev

Plugin Development for Claude Code & OpenCode

Critical Importance

Creating high-quality plugins is critical to your development workflow's long-term success. Plugins are used repeatedly by yourself and others. Design flaws, poor documentation, or broken functionality compound over time and across users. A well-designed plugin becomes a trusted tool used daily; a poorly designed plugin becomes abandoned technical debt. Invest time in architecture, testing, and documentation—the returns multiply across all future uses.

Systematic Approach

** approach plugin development systematically.** Plugins require careful planning: understand the problem, design the API, implement incrementally, test thoroughly, and document comprehensively. Don't rush to code—clarify requirements, define interfaces, and consider edge cases first. Build iteratively, validate frequently, and refactor continuously. Every design decision impacts maintainability and extensibility.

The Challenge

The create a plugin that balances specificity with flexibility perfectly, but if you can:

• Your plugin will be a joy to use and extend
• Others will build on top of your work
• The plugin will remain useful as needs evolve
• You'll establish patterns for future plugin development

The challenge is designing plugins that solve specific problems while staying flexible enough for future use cases. Can you create focused, opinionated tools that don't paint yourself into corners?

Plugin Confidence Assessment

After completing or reviewing plugin development, rate your confidence from 0.0 to 1.0:

• 0.8-1.0: Plugin well-architected, fully tested, thoroughly documented, follows platform conventions
• 0.5-0.8: Plugin functional but missing some tests or documentation, some technical debt
• 0.2-0.5: Plugin works but design unclear, minimal testing, poor documentation
• 0.0-0.2: Plugin incomplete or broken, unclear purpose, significant rework needed

Identify uncertainty areas: Is the plugin's purpose clear? Are there edge cases unhandled? Will the plugin work as requirements change? What's the maintenance burden?

Overview

The ai-eng-system supports extension development for both Claude Code and OpenCode through a unified content system with automated transformation. Understanding this system enables creating well-organized, maintainable extensions that integrate seamlessly with both platforms.

Extension Types

Type	Claude Code	OpenCode	Shared Format
Commands	✅ YAML frontmatter	✅ Table format	YAML frontmatter
Agents	✅ YAML frontmatter	✅ Table format	YAML frontmatter
Skills	✅ Same format	✅ Same format	SKILL.md
Hooks	✅ hooks.json	✅ Plugin events	Platform-specific
Custom Tools	❌ (use MCP)	✅ tool() helper	OpenCode only
MCP Servers	✅ .mcp.json	✅ Same format	Same format

Development Approaches

1. Canonical Development (Recommended)

Create content in content/ directory, let build.ts transform to platform formats:

content/
├── commands/my-command.md  → dist/.claude-plugin/commands/
│                           → dist/.opencode/command/ai-eng/
└── agents/my-agent.md      → dist/.claude-plugin/agents/
                            → dist/.opencode/agent/ai-eng/

2. Platform-Specific Development

Create directly in platform directories:

• Claude Code: .claude/commands/, .claude-plugin/
• OpenCode: .opencode/command/, .opencode/agent/

3. Global vs Project-Local

Location	Claude Code	OpenCode
Project	.claude/	.opencode/
Global	~/.claude/	~/.config/opencode/

Lifecycle workflows

Plan architecture (new or refactor)

1. Clarify goal, users, and outcomes.
2. Pick the smallest component set: rules, skills, agents, commands, hooks, mcpServers.
3. Propose directory layout and manifest. Default local path: ~/.cursor/plugins/local/<plugin-name>/.
4. Flag discoverability and metadata risks early.
5. Return an implementation checklist.

Scaffold a new plugin

Inputs: kebab-case name, purpose, component set, single-plugin vs marketplace repo.

Steps:

1. Validate name (lowercase kebab-case, alphanumeric ends).
2. Create target dir (default ~/.cursor/plugins/local/<plugin-name>/ unless user specifies otherwise).
3. Add .cursor-plugin/plugin.json, README.md, LICENSE, optional CHANGELOG.md.
4. Populate manifest (name required; recommend version, description, author, license, keywords).
5. Add components with valid frontmatter (rules .mdc, skills skills/*/SKILL.md, agents/commands markdown).
6. For marketplace repos, register in .cursor-plugin/marketplace.json with relative source.
7. Verify paths are relative—no absolute paths or parent traversal.

Guardrails: one focused use case; concise actionable text; no references to missing files.

Pre-submission review

Before release or marketplace submission:

1. Manifest: valid plugin.json, kebab-case name, coherent metadata.
2. Discoverability: skills, rules, agents, commands, hooks, MCP in expected paths.
3. Frontmatter: every skill/rule/agent/command has required fields.
4. Repo integration: marketplace entry if applicable; unique names; README covers purpose and install.
5. Report: pass/fail by section, prioritized fixes, submission recommendation.

Quick Reference

Command Frontmatter

Canonical (content/):

---
name: my-command
description: What this command does
agent: build           # Optional: which agent handles this
subtask: true          # Optional: run as subtask
temperature: 0.3      # Optional: temperature
tools:                 # Optional: tool restrictions
  read: true
  write: true
---

Claude Code Output: Same format (YAML frontmatter)

OpenCode Output: YAML frontmatter with OpenCode-compatible fields

---
description: Description here
agent: build
---

Agent Frontmatter

Canonical (content/):

---
name: my-agent
description: Use this agent when... <example>...</example>
mode: subagent
color: cyan
temperature: 0.3
tools:
  read: true
  write: true
---

Claude Code Output: Same format (YAML frontmatter)

OpenCode Output: YAML frontmatter with OpenCode-compatible fields

---
description: Description here
mode: subagent
---

Skill Structure

Both platforms use identical format:

skill-name/
├── SKILL.md (required)
│   ├── YAML frontmatter (name, description)
│   └── Markdown body (1,000-3,000 words)
└── Bundled Resources (optional)
    ├── references/       # Detailed documentation
    ├── examples/         # Working code
    └── scripts/          # Utility scripts

OpenCode Custom Tools

Use TypeScript with tool() helper:

import { tool } from "@opencode-ai/plugin"

export default tool({
  description: "Tool description",
  args: {
    param: tool.schema.string().describe("Parameter description"),
  },
  async execute(args, context) {
    // Tool implementation
    return result
  },
})

Directory Locations

For Development in ai-eng-system

ai-eng-system/
├── content/
│   ├── commands/              # Add new commands here
│   └── agents/                # Add new agents here
├── skills/
│   └── plugin-dev/           # This skill
└── build.ts                   # Transforms to both platforms

For User Projects

Project-local:

• Claude Code: .claude/commands/, .claude-plugin/
• OpenCode: .opencode/command/, .opencode/agent/

Global:

• Claude Code: ~/.claude/commands/, ~/.claude-plugin/
• OpenCode: ~/.config/opencode/command/, ~/.config/opencode/agent/

Platform-Specific Features

Claude Code

Components:

• Commands with YAML frontmatter
• Agents with YAML frontmatter
• Skills with SKILL.md format
• Hooks via hooks/hooks.json
• MCP servers via .mcp.json

Manifest: .claude-plugin/plugin.json

{
  "name": "plugin-name",
  "version": "1.0.0",
  "description": "Brief description",
  "commands": ["./commands/*"],
  "mcpServers": "./.mcp.json"
}

OpenCode

Components:

• Commands with table format
• Agents with table format
• Skills via opencode-skills plugin
• Custom tools with TypeScript
• Plugin events via TypeScript

Plugin: .opencode/plugin/plugin.ts

import { Plugin } from "@opencode-ai/plugin"

export default (async ({ client, project, directory, worktree, $ }) => {
  return {
    // Plugin hooks here
  }
}) satisfies Plugin

Development Workflow

1. Create Component

Use plugin-dev commands:

• /ai-eng/create-agent - Create new agent
• /ai-eng/create-command - Create new command
• /ai-eng/create-skill - Create new skill
• /ai-eng/create-tool - Create new custom tool

2. Build

cd ai-eng-system
bun run build              # Build all platforms
bun run build --watch        # Watch mode
bun run build --validate      # Validate content

3. Test

Claude Code:

claude plugin add https://github.com/v1truv1us/ai-eng-system

OpenCode:

# Project-local
./setup.sh

# Global
./setup-global.sh

Best Practices

Content Quality

• Use third-person in skill descriptions
• Write commands/agents FOR Claude, not to user
• Include specific trigger phrases
• Follow progressive disclosure for skills

File Organization

• One component per file
• Clear naming conventions (kebab-case)
• Proper frontmatter validation

Cross-Platform Compatibility

• Use canonical format in content/
• Test build output for both platforms
• Document platform differences

Security

• No hardcoded credentials
• Use HTTPS/WSS for external connections
• Validate user inputs
• Follow principle of least privilege

Additional Resources

References

• references/claude-code-plugins.md - Claude Code specifics
• references/opencode-plugins.md - OpenCode specifics
• references/command-format.md - Command syntax guide
• references/agent-format.md - Agent configuration guide
• references/skill-format.md - Skills specification
• references/opencode-tools.md - Custom tool development

Anti-Rationalization Table

Excuse	Counter
"I'll test the plugin after I finish all features"	Test incrementally. Untested plugins accumulate bugs that are hard to trace.
"The canonical format is optional"	Canonical format ensures cross-platform compatibility. Use it or break one platform.
"I don't need to document platform differences"	Platform differences are where plugins break. Document them or users will hit them.
"Security is not a concern for internal plugins"	Internal plugins run with the same privileges as external ones. Validate inputs always.
"The frontmatter fields are self-explanatory"	Missing or malformed frontmatter causes silent failures. Validate before deploying.

Examples

Study existing components in ai-eng-system:

• content/commands/plan.md - Command structure
• content/agents/architect-advisor.md - Agent structure
• skills/prompting/incentive-prompting/SKILL.md - Skill structure

Troubleshooting

Build Issues

• Run bun run build --validate to check content
• Check file permissions in output directories
• Verify YAML frontmatter syntax

Platform Testing

• Test commands in both Claude Code and OpenCode
• Verify agents trigger correctly
• Check skills load via opencode-skills plugin

Common Errors

• Missing required frontmatter fields
• Incorrect directory structure
• Invalid YAML syntax
• Wrong file permissions

Integration with AI Engineering System

The plugin-dev system integrates seamlessly with existing ai-eng-system components:

Existing Commands

• /ai-eng/plan - Implementation planning
• /ai-eng/review - Code review
• /ai-eng/work - Task execution

Existing Agents

• ai-eng/architect-advisor - Architecture guidance
• ai-eng/frontend-reviewer - Frontend review
• ai-eng/seo-specialist - SEO optimization

Plugin-Dev Commands

• /ai-eng/create-plugin - Full plugin development workflow
• /ai-eng/create-agent - Quick agent creation
• /ai-eng/create-command - Quick command creation
• /ai-eng/create-skill - Quick skill creation
• /ai-eng/create-tool - Quick tool creation

All use the same quality standards and research-backed prompting techniques.