writing-plugins

Writing Plugins

Scope: covers plugin design and architecture. The examples use the Claude Code layout (.claude-plugin/plugin.json + auto-discovered commands/, agents/, skills/, hooks/). The same component-selection and architecture reasoning maps to the other tools — only the manifest path and packaging differ:

• Codex CLI: .codex-plugin/plugin.json manifest + .agents/plugins/marketplace.json; skills live at .agents/skills/. See [[nlpm:conventions-codex]].
• Antigravity: gemini-extension.json (becoming "Antigravity plugins"); skills at .agent/skills/. See [[nlpm:conventions-antigravity]].
• Cross-tool skills: a SKILL.md collection with no plugin wrapper installs into any tool via npx skills add. See [[writing-skills]].

For individual component authoring, see [[writing-skills]], [[writing-agents]], [[writing-hooks]], [[writing-rules]].

1. Plugin = Commands + Agents + Skills + Hooks

A plugin is a collection of NL artifacts that work together. Before writing anything, decide which components you need.

Component Selection Guide

User need	Component	Example
User runs a slash command	Command	/nlpm:score path/to/file.md
AI works autonomously on a task	Agent	Security scanner dispatched by a command
Domain knowledge for agents/Claude	Skill	SKILL.md with patterns and decision tables
Something must happen automatically on events	Hook	Lint-on-save, block force push
External service integration	MCP server (.mcp.json)	GitHub API, Slack notifications

Minimum Viable Plugin

The smallest useful plugin has one component:

my-plugin/
  .claude-plugin/
    plugin.json
  commands/
    do-thing.md

Don't add agents, skills, or hooks until you need them. Each component adds maintenance burden.

2. Architecture Patterns

Pattern: Single Command (simplest)

Command does everything itself. No agents, no orchestration.

Command receives input --> processes --> outputs result

Use when: task is simple, deterministic, single-step. Example: loc-guardian /scan -- counts lines, checks limits, reports.

File structure:

my-plugin/
  .claude-plugin/plugin.json
  commands/do-thing.md

Pattern: Command + Agent

Command parses input and dispatches one agent for heavy work.

Command parses input --> dispatches agent --> formats output

Use when: task requires AI judgment but has a clear entry point. Example: /nlpm:score dispatches a scorer agent.

File structure:

my-plugin/
  .claude-plugin/plugin.json
  commands/analyze.md
  agents/analyzer.md

Pattern: Command + Multiple Agents (parallel)

Command dispatches 2–6 agents in parallel, synthesizes results.

Command --> agent-1 (security)
        --> agent-2 (performance)    --> synthesize --> output
        --> agent-3 (architecture)

Use when: multiple independent analyses of the same input. Example: grill dispatches 6 review agents in parallel.

File structure:

my-plugin/
  .claude-plugin/plugin.json
  commands/review.md
  agents/
    security-agent.md
    performance-agent.md
    architecture-agent.md

Pattern: Command + Agent Pipeline (sequential)

Each agent feeds into the next. Stages have different model requirements.

Command --> parse (haiku) --> analyze (sonnet) --> QC (sonnet) --> output

Use when: multi-phase processing where each phase depends on the previous. Example: reading-assistant's 4-phase pipeline.

File structure:

my-plugin/
  .claude-plugin/plugin.json
  commands/process.md
  agents/
    parser.md       # haiku
    analyzer.md     # sonnet
    qc-agent.md     # sonnet

Pattern: Hooks Only (no commands)

Plugin enforces policy silently via hooks. No user-facing commands.

Event fires --> hook checks --> allow/deny/advise

Use when: enforcement should be automatic, not user-initiated. Example: tdd-guardian's pre-commit quality gate.

File structure:

my-plugin/
  .claude-plugin/plugin.json
  hooks/hooks.json
  scripts/check.sh

Pattern Selection Matrix

Question	Yes -->	No -->
Does the user explicitly trigger it?	Needs a command	Hooks only
Does it require AI judgment?	Needs agents	Command-only or hooks
Are there independent sub-analyses?	Parallel agents	Sequential or single agent
Does each step depend on the previous?	Sequential pipeline	Parallel or single
Should it run automatically on events?	Add hooks	Commands only
Does Claude need domain knowledge?	Add skills	No skills needed

3. The plugin.json Manifest

Required Fields

Only name is strictly required:

{
  "name": "my-plugin"
}

Recommended Fields

Ship with these for discoverability and marketplace listing:

{
  "name": "my-plugin",
  "version": "0.1.0",
  "description": "What this plugin does in one sentence",
  "author": { "name": "your-name" },
  "license": "MIT",
  "keywords": ["relevant", "search", "terms"],
  "category": "developer-tools"
}

Field Reference

Field	Purpose	Example
name	Unique identifier, used in slash commands	"nlpm"
version	Semver, used by marketplace and update checks	"0.1.0"
description	One-line summary for marketplace listing	"NL programming quality tools"
author.name	Creator attribution	"xiaolai"
license	Open source license	"MIT"
keywords	Search terms for marketplace discovery	["linter", "quality"]
category	Marketplace category	"developer-tools"

4. File Structure

Full Plugin Layout

my-plugin/
  .claude-plugin/
    plugin.json              # manifest (required)
    marketplace.json         # for marketplace publishing
  commands/                  # auto-discovered by Claude Code
    do-thing.md              # user-invocable: /plugin:do-thing
    advanced-thing.md
    shared/                  # non-invocable partials
      common-logic.md        # user-invocable: false
      format-output.md
  agents/                    # auto-discovered
    worker.md
    reviewer.md
  skills/                    # auto-discovered
    my-plugin/
      domain-knowledge/
        SKILL.md
      advanced-topic/
        SKILL.md
        references/
          deep-dive.md
  hooks/
    hooks.json               # hook definitions
  scripts/                   # hook scripts, utilities
    check.sh
    validate.sh
  CLAUDE.md                  # architecture guide (for Claude)
  README.md                  # user documentation (for humans)
  LICENSE

Directory Conventions

Directory	Auto-discovered?	What goes here
.claude-plugin/	Yes (manifest)	Plugin metadata
commands/	Yes	Slash command definitions
commands/shared/	Yes (but not invocable)	Shared command logic
agents/	Yes	Agent definitions
skills/	Yes	Domain knowledge
hooks/	Yes (hooks.json)	Event hooks
scripts/	No	Shell scripts, utilities

Naming Conventions

Component	File naming	Example
Commands	kebab-case, descriptive verb	scan-files.md, generate-report.md
Agents	kebab-case, role-noun	security-reviewer.md, parser.md
Skills	kebab-case directory, always SKILL.md	skills/my-plugin/react-patterns/SKILL.md
Hook scripts	kebab-case, descriptive	check-loc.sh, validate-config.sh

5. Versioning

Semver Rules

Change type	Bump	Example
Bug fixes, typo corrections, penalty adjustments	Patch: 0.1.0 -> 0.1.1	Fix scoring formula
New commands, new agents, new features	Minor: 0.1.0 -> 0.2.0	Add /plugin:export command
Breaking changes (renamed commands, removed features)	Major: 0.1.0 -> 1.0.0	Rename /scan to /analyze

Four-Place Update

When bumping version, update in four places:

Location	File	Field
1. Plugin manifest	.claude-plugin/plugin.json	version
2. Plugin marketplace	.claude-plugin/marketplace.json	version in the plugin's entry
3. Central marketplace manifest	~/.claude/plugins/marketplaces/xiaolai/.claude-plugin/marketplace.json	version
4. Central marketplace README	~/.claude/plugins/marketplaces/xiaolai/README.md	Version in the table

Order: push plugin repo first, then update central marketplace. The marketplace points to the repo -- if the repo isn't updated yet, users pull stale code.

6. CLAUDE.md for Plugins

Your plugin's CLAUDE.md is for Claude (the AI), not the user. It tells Claude how the plugin's components relate to each other.

What to Include

# my-plugin

## Architecture
Brief description of what the plugin does and how components interact.

## Components

### Commands
| Command | Purpose |
|---------|---------|
| /plugin:scan | Discovers and inventories files |
| /plugin:fix | Auto-fixes issues found by scan |

### Agents
| Agent | Model | Role |
|-------|-------|------|
| scanner | haiku | Mechanical file discovery |
| fixer | sonnet | AI-powered fix generation |

### Conventions
- All agents output findings in severity-tagged format
- Scanner runs before fixer (sequential dependency)
- Hook scripts use fail-open pattern

What NOT to Include

• Installation instructions (those go in README.md)
• User-facing documentation (README.md)
• Changelog (CHANGELOG.md or git history)
• Contributing guidelines (CONTRIBUTING.md)

7. Shared Partials

Extract repeated logic into commands/shared/*.md with user-invocable: false.

Partial Frontmatter

---
user-invocable: false
description: "Shared config loading logic"
---

Good Candidates for Extraction

Partial	Content	When to extract
shared/load-config.md	Read and validate config file	3+ commands need config
shared/discover-files.md	Find target files by pattern	3+ commands scan files
shared/validate-prereqs.md	Check tool availability	2+ commands need same tools
shared/format-report.md	Report header, footer, severity format	3+ commands output reports

When NOT to Extract

• Logic used by only 1 command (premature abstraction)
• Simple logic under 10 lines (duplication is fine)
• Logic that differs slightly between commands (forced generalization adds complexity)

8. Testing Your Plugin

Pre-Publish Checklist

Check	How	Pass criteria
Structure validation	claude plugin validate /path/to/plugin	No errors
Command: no args	Run each command with no arguments	Helpful error or usage message
Command: normal args	Run each command with typical input	Correct output
Command: edge cases	Empty files, huge files, missing files	Graceful error handling
Agent triggering	Try queries that should and shouldn't trigger	Correct dispatch decisions
Hook scripts	chmod +x check, run with test JSON	Valid JSON output
Hook fail-open	Kill script mid-execution	Action is allowed

Testing Agent Triggers

For each agent, test with 3 types of queries:

Query type	Expected	Example
Direct match	Agent triggers	"Scan this code for security issues"
Adjacent topic	Agent may or may not trigger	"Is this code okay?"
Unrelated	Agent does NOT trigger	"What's the weather?"

9. Marketplace Publishing

1. Push your plugin repo to GitHub
2. Add entry to central marketplace.json (name, source, description, version, author, license, keywords, category)
3. Add row to central marketplace README.md version table
4. Commit and push the central marketplace repo
5. Verify: claude plugin install my-plugin@xiaolai --scope project

Pre-publish checks: claude plugin validate ., verify no hardcoded paths (grep -r '/Users/' commands/ agents/ hooks/), verify all scripts executable, verify version matches in all 4 locations.

10. Common Mistakes

Mistake	Impact	Fix
Commands that do too much	Hard to maintain, unreliable	Split into focused commands
Agents without examples	40% trigger accuracy	Add 2-3 specific scenario examples
Skills over 500 lines	Context bloat, slow loading	Extract to references/ subdirectory
Hooks that block without explanation	Frustrating UX	Always include permissionDecisionReason
No CLAUDE.md	Claude doesn't understand plugin architecture	Add architecture overview
README documents internals	Users confused by implementation details	README = user guide, CLAUDE.md = internals
Hardcoded paths	Breaks on other machines	Use ${CLAUDE_PLUGIN_ROOT} everywhere
No error handling in commands	Silent failures	Add explicit error cases
Version not updated in all 4 places	Marketplace shows wrong version	Use the four-place update checklist
Premature extraction into shared/	Over-abstracted, harder to understand	Extract only when 3+ consumers exist