Plugin Creator

Claude Code Plugin Creator — Agentic Orchestration Workflow

This skill orchestrates specialized agents through a comprehensive plugin creation workflow. The orchestrator (you) delegates to sub-agents for research, discovery, validation, and implementation — never performing these tasks directly.

Workflow Diagram: See references/workflow-diagram.md for mermaid flowcharts of the complete plugin creation flow.

Orchestration Principles

The orchestrator delegates — it does not execute.

flowchart TD
    Start([Task arrives]) --> Q1{User names a specific agent?}
    Q1 -->|Yes — explicit instruction| Override["Use exactly the agent named<br>Explicit instruction overrides all routing tables"]
    Q1 -->|No| Q2{Task type?}
    Q2 -->|Domain research or reasoning| CG["plugin-assessor agent<br>subagent_type='plugin-creator:plugin-assessor'"]
    Q2 -->|Verbatim file/keyword retrieval only| Explore["Explore agent — Haiku-based<br>⚠️ retrieval only, never reasoning"]
    Q2 -->|Official docs fetch or analysis| GP["general-purpose agent<br>subagent_type='general-purpose'"]
    Q2 -->|Schema validation| Scripts[Validation scripts]
    Q2 -->|Quality review| Assessor["plugin-assessor agent"]
    Q2 -->|Documentation writing| Docs["plugin-docs-writer agent"]
    Override --> Done([Delegate])
    CG --> Done
    Explore --> Done
    GP --> Done
    Scripts --> Done
    Assessor --> Done
    Docs --> Done

Explore agent constraint (hard rule from CLAUDE.md): Explore uses Haiku internally. Validated failure rate ~50% for reasoning tasks (2026-02-02). Use Explore ONLY for verbatim retrieval — exact file contents, directory listings, keyword searches with no interpretation required. Any task requiring analysis, comparison, or judgment goes to plugin-creator:plugin-assessor or general-purpose.

How to spawn agents:

• Agent tool (subagent_type=...) — for focused, isolated work where only the result matters
• TeamCreate — for multi-agent coordination where agents need to communicate (see agent-orchestration:parallel-work "Persistent teams")

Why delegation matters:

1. Sub-agents have focused context and specialized prompts
2. Delegation creates audit trails of verified information
3. Prevents hallucination by requiring source verification
4. Enables parallel work and thoroughness

Component Selection Guidance

When deciding which component type to create (skill, agent, hook, MCP server, or command), use /plugin-creator:component-patterns for the complete decision framework covering component lifecycle, discovery and activation phases, and organization patterns.

Artifact System

Maintain structured artifacts for crash recovery, audit trails, and context management.

Create a work directory for each plugin project:

.plugin-creator/plans/{plugin-name}/
├── PROJECT.md                # Vision and goals (always loaded)
├── REQUIREMENTS.md           # Scoped deliverables
├── STATE.md                  # Decisions, blockers, current position
├── discuss-CONTEXT.md        # User preferences captured in discussion
├── research-FINDINGS.md      # 4-way parallel research results
├── design-PLAN.md            # Architecture with XML task specs
├── validation-REPORT.md      # Multi-layer verification results
└── SUMMARY.md                # Completion record

STATE.md format (persists across sessions):

# Plugin State: {plugin-name}
Last Updated: {ISO timestamp}

## Decisions Made
- {decision}: {rationale}

## Current Position
- Phase: {current phase}
- Status: IN_PROGRESS | BLOCKED | COMPLETE

## Blockers
- {blocker}: {what's needed}

## Deviations from Plan
- {change}: {why}

Recovery: Read STATE.md to restore context after session crash.

Parallel Agent Spawning

Spawn independent agents simultaneously to maximize throughput.

4-Way Parallel Research Pattern:

# Spawn all four researchers in a single message:

Agent(subagent_type="plugin-creator:plugin-assessor", prompt="EXISTING PLUGINS: Search plugins/ and ~/.claude/skills/ for similar functionality...")
Agent(subagent_type="plugin-creator:plugin-assessor", prompt="CLAUDE CODE FEATURES: What plugin capabilities exist? Dynamic context, hooks, MCP, LSP...")
Agent(subagent_type="plugin-creator:plugin-assessor", prompt="ARCHITECTURE PATTERNS: How do well-structured plugins organize skills, agents, references...")
Agent(subagent_type="general-purpose", prompt="PITFALLS: Fetch official docs, identify common mistakes, schema gotchas...")

All four run concurrently. Merge results into research-FINDINGS.md before planning.

Parallelization opportunities:

Sequential requirements:

• Discussion phase captures preferences BEFORE research
• Design phase depends on merged research findings
• Implementation depends on approved design
• Plan checker must pass BEFORE execution

Plugin Runtime Mechanics

Read this before any plugin work begins.

A CLAUDE.md inside a plugin directory informs Claude when it is doing development work inside that plugin directory. It has zero effect on plugin users because Claude runs in the user's project directory, not the plugin source directory.

flowchart TD
    Q{Where do I put this content?} --> A1{Who needs to see it?}
    A1 -->|"Plugin users — enforce behavior,<br>deliver instructions, add tools"| Runtime["Put it in runtime-injected components"]
    A1 -->|"Claude during development — understand design,<br>conventions, architecture"| Dev["Put it in CLAUDE.md<br>inside the plugin directory"]

    Runtime --> R1["SKILL.md — loaded when skill is invoked<br>Instructions land in the user's session"]
    Runtime --> R2["hooks.json — runs on session/tool events<br>Enforces rules without user action"]
    Runtime --> R3["Agent definitions — loaded when agent is used<br>Delivers specialized instructions"]
    Runtime --> R4["MCP servers — tools available throughout session"]

    Dev --> D1["Loaded only when Claude is working in<br>the plugin directory during a development session<br>Invisible to plugin consumers"]

The mistake this corrects: Placing behavioral rules in rules/CLAUDE.md or CLAUDE.md inside a plugin directory does not enforce those rules on users. To enforce behavior on users, the content must be in SKILL.md or delivered via a hook.

Entry Gate: Collect Intent Before Starting

$ARGUMENTS

If the plugin concept above or the conversation context indicates anything other than creating a new plugin from scratch (for example: improving, fixing, validating, refactoring, auditing, adding a component to, or working on an existing plugin), load plugin-lifecycle instead: Skill(skill="plugin-creator:plugin-lifecycle"). Do not proceed further in this skill.

If no plugin concept is provided and the intent is not yet clear, STOP and ask: "Are you creating a new plugin from scratch, or working on an existing one?" Route to plugin-lifecycle for existing; continue here for new.

If creating a new plugin and no concept is provided, STOP and ask: "What plugin do you want to create? Describe it in 1-2 sentences." Wait for the answer before proceeding to Phase 0.

Phase 0: RT-ICA Prerequisite Check

STOP. Before creating any plugin, perform RT-ICA assessment.

Invoke the rt-ica skill to verify prerequisites:

RT-ICA SUMMARY
==============

Goal:
- Create a Claude Code plugin for [purpose]

Success Output:
- Functional plugin that [specific outcome]

Conditions (reverse prerequisites):
1. Purpose clarity | Requires: Clear problem statement | Why: Determines plugin scope
2. Target users | Requires: Who will use this | Why: Shapes UX decisions
3. Component selection | Requires: Skills vs Agents vs Hooks | Why: Architecture
4. Existing solutions | Requires: Check for similar plugins | Why: Avoid duplication
5. Source material | Requires: Documentation/APIs to encode | Why: Content accuracy
6. Verification method | Requires: How to test the plugin works | Why: Quality gate

Verification:
- [Check each condition: AVAILABLE / DERIVABLE / MISSING]

Decision:
- [APPROVED / BLOCKED]

IF BLOCKED: Request missing information before proceeding.

IF APPROVED: Continue to Discussion phase.

Phase 0.5: Discussion (Capture Preferences)

BEFORE research, identify gray areas and capture user preferences.

Ask targeted questions to eliminate ambiguity:

For skill-focused plugins:

• Activation triggers: When should Claude auto-load vs user-invoke?
• Tool restrictions: Full access or limited tools?
• Output format: Verbose explanations or terse instructions?
• Reference structure: Inline content or progressive disclosure?

For agent-focused plugins:

• Delegation scope: What tasks should agents handle?
• Return format: Summaries or detailed reports?
• Error handling: Retry, escalate, or fail fast?
• For agent frontmatter schema (tools, model, permissionMode, memory, hooks, etc.), load /plugin-creator:claude-subagent-reference.

For hook-focused plugins:

• Trigger events: Which tool/session events matter?
• Hook type: Command, prompt, or agent verification?
• Timeout handling: Fail silently or block?

Save preferences to discuss-CONTEXT.md:

# Plugin Discussion: {plugin-name}
Date: {ISO timestamp}

## Scope Decisions
- {question}: {user preference}

## UX Preferences
- Invocation: {user-invoked | model-invoked | both}
- Verbosity: {terse | balanced | verbose}

## Technical Choices
- {choice}: {preference with rationale}

These preferences guide all subsequent research and planning.

Phase 1: Research (4-Way Parallel)

Spawn 4 parallel researchers in a single message. Each investigates a different domain:

# Launch all four simultaneously:

Agent(subagent_type="plugin-creator:plugin-assessor", prompt="
RESEARCHER 1: EXISTING SOLUTIONS
Search for plugins/skills similar to {plugin-name}:
- plugins/ directory
- ~/.claude/skills/
- GitHub repos with Claude Code plugins
REPORT: What exists, gaps to fill, patterns to follow/avoid
Write findings to .plugin-creator/plans/{plugin-name}/research-1-existing.md")

Agent(subagent_type="plugin-creator:plugin-assessor", prompt="
RESEARCHER 2: CLAUDE CODE FEATURES
What capabilities should this plugin use?
- Dynamic context injection (!command)
- Subagent execution (context: fork)
- Hooks (which events?)
- MCP/LSP integration opportunities
REPORT: Recommended features with rationale
Write findings to .plugin-creator/plans/{plugin-name}/research-2-features.md")

Agent(subagent_type="plugin-creator:plugin-assessor", prompt="
RESEARCHER 3: ARCHITECTURE PATTERNS
How do well-structured plugins organize?
- Skill directory structure
- Reference file patterns
- Agent definitions
- Hook configurations
REPORT: Recommended structure based on similar plugins
Write findings to .plugin-creator/plans/{plugin-name}/research-3-architecture.md")

Agent(subagent_type="general-purpose", prompt="
RESEARCHER 4: PITFALLS & OFFICIAL DOCS
Fetch https://code.claude.com/docs/en/plugins-reference.md
Fetch https://code.claude.com/docs/en/skills.md
IDENTIFY:
- Schema requirements (comma-separated strings NOT arrays)
- Common mistakes
- Deprecations or new features
REPORT: Gotchas to avoid, schema requirements
Write findings to .plugin-creator/plans/{plugin-name}/research-4-pitfalls.md")

Merge all 4 reports into research-FINDINGS.md before proceeding to Design.

Research Findings Format

# Research Findings: {plugin-name}
Date: {ISO timestamp}

## 1. Existing Solutions
{Researcher 1 findings}

## 2. Recommended Features
{Researcher 2 findings}

## 3. Architecture Patterns
{Researcher 3 findings}

## 4. Pitfalls & Requirements
{Researcher 4 findings}

## Synthesis
- Key insights: {combined learnings}
- Recommended approach: {synthesis}

Phase 2: Design (Plan + Verify Loop)

Design phase uses a PLAN → CHECK → ITERATE loop until verification passes.

2a. Generate Plan with XML Task Specs

Delegate to Plan agent:

Agent(
  agent="Plan",
  prompt="Design plugin: {plugin-name}

  INPUTS:
  - User preferences: {from discuss-CONTEXT.md}
  - Research findings: {from research-FINDINGS.md}

  OUTPUT: XML task specifications for atomic implementation:

  <task id='1' type='auto'>
    <name>Create plugin.json manifest</name>
    <files>.claude-plugin/plugin.json</files>
    <action>Create manifest with name, version, description. Do NOT add a skills field — skills under ./skills/ are auto-discovered by Claude Code (Mode A). Add a skills field only when explicitly opting into manual allowlist mode (Mode B).</action>
    <verify>jq '.name' .claude-plugin/plugin.json returns plugin name</verify>
    <done>Valid plugin.json exists with all required fields</done>
  </task>

  <task id='2' type='auto'>
    <name>Create main SKILL.md</name>
    <files>skills/{skill-name}/SKILL.md</files>
    <action>Create skill with frontmatter and core instructions</action>
    <verify>grep -q '^---' skills/{skill-name}/SKILL.md</verify>
    <done>SKILL.md has valid frontmatter and passes token-count validation</done>
  </task>

  Generate 2-5 atomic tasks. Each task must have:
  - Single responsibility
  - Testable <verify> command
  - Clear <done> criteria"
)

2b. Plan Checker Verification

BEFORE execution, verify the plan achieves goals:

Agent(
  agent="general-purpose",
  prompt="PLAN CHECKER: Verify this plan achieves the plugin goals.

  PLAN: {generated XML tasks}
  REQUIREMENTS: {from discuss-CONTEXT.md}
  RESEARCH: {key findings}

  VERIFY:
  1. Do tasks cover all required components?
  2. Are tasks truly atomic (single responsibility)?
  3. Are <verify> commands actually testable?
  4. Are there gaps between tasks?
  5. Does sequence respect dependencies?

  OUTPUT:
  - PASS: Plan is ready for execution
  - FAIL: {specific issues to fix}

  If FAIL, return to planner with feedback."
)

Loop until plan checker returns PASS.

2c. Save Approved Plan

Save to design-PLAN.md:

# Design Plan: {plugin-name}
Date: {ISO timestamp}
Status: APPROVED

## Tasks

<task id='1'>...</task>
<task id='2'>...</task>

## Verification
Plan checker: PASS
Reviewer: {agent ID}

Phase 3: Implementation (Atomic Execution)

Execute each XML task atomically with per-task commits.

3a. Task Execution Pattern

For each <task> in the approved plan:

Agent(
  agent="general-purpose",
  prompt="EXECUTOR: Implement this single task.

  <task id='{N}'>
    <name>{task name}</name>
    <files>{target files}</files>
    <action>{implementation instructions}</action>
    <verify>{test command}</verify>
    <done>{success criteria}</done>
  </task>

  CONTEXT:
  - User preferences: {from discuss-CONTEXT.md}
  - Research findings: {relevant sections}

  EXECUTE:
  1. Implement the <action>
  2. Run the <verify> command
  3. Confirm <done> criteria met

  OUTPUT:
  - Files created/modified
  - Verification result: PASS/FAIL
  - If FAIL: what went wrong"
)

3b. Atomic Git Commits

Each task gets its own commit immediately after completion:

git add {files from task}
git commit -m "task-{N}: {task name}"

Benefits:

• git bisect locates exact failing task
• Individual tasks revertable
• Clear history for debugging

3c. Parallel vs Sequential Execution

Independent tasks (no shared files): Execute in parallel

# Tasks 1, 2, 3 have no dependencies — spawn all:
Agent(prompt="EXECUTOR: task 1...")
Agent(prompt="EXECUTOR: task 2...")
Agent(prompt="EXECUTOR: task 3...")

Dependent tasks (task 2 needs task 1's output): Execute sequentially

3d. Scaffolding Script Option

For simple plugins, use the scaffolding script:

uv run scripts/create_plugin.py create my-plugin -d "Description" -s my-skill -o ./plugins

The script self-validates created files.

3e. Advanced Features

See references/advanced-features.md for:

• Dynamic context injection (!command`` syntax)
• String substitutions ($ARGUMENTS, ${CLAUDE_SESSION_ID}, ${CLAUDE_PLUGIN_ROOT})
• Running skills in subagents (context: fork)
• Visual output via bundled scripts
• Hook configuration (all event types with examples)
• MCP and LSP server integration
• Plugin caching behavior and path rules
• Skill invocation control (disable-model-invocation, user-invocable)
• Extended thinking (ultrathink)

Manual Implementation Structure

Directory structure:

my-plugin/
├── .claude-plugin/           # REQUIRED: metadata directory
│   └── plugin.json          # REQUIRED: only file in .claude-plugin/
├── agents/                   # Optional: agent definitions (.md)
├── skills/                   # Optional: skill directories
│   └── my-skill/
│       ├── SKILL.md
│       └── references/      # Optional: detailed reference docs
├── hooks/                    # Optional: hook configurations
│   └── hooks.json
├── .mcp.json                # Optional: MCP server definitions
├── scripts/                 # Optional: helper scripts
├── LICENSE
└── README.md

Critical Rules:

1. .claude-plugin/ contains ONLY plugin.json
2. All components go at plugin root, NOT inside .claude-plugin/
3. Commands in plugins are deprecated — use skills instead

plugin.json Schema

{
  "name": "my-plugin",
  "version": "1.0.0",
  "description": "What this plugin does and trigger keywords",
  "author": {
    "name": "Your Name",
    "email": "you@example.com"
  },
  "repository": "https://github.com/you/my-plugin",
  "license": "MIT",
  "keywords": ["keyword1", "keyword2"],
  "skills": ["./skills/my-skill"]
}

Source: https://code.claude.com/docs/en/plugins-reference.md#plugin-manifest-schema

The settings field (inline {"agent": "agent-name"}) activates a plugin agent as the main thread, applying its system prompt and tool restrictions as the default behavior. SOURCE: https://code.claude.com/docs/en/plugins.md (accessed 2026-03-07)

SKILL.md Frontmatter

---
description: 'Detailed description including trigger keywords. Use when [situation].'
allowed-tools: Read, Grep, Glob
---

CRITICAL: YAML frontmatter fields like allowed-tools MUST be comma-separated strings, NOT arrays.

Source: https://code.claude.com/docs/en/skills.md

Phase 4: Validation (Multi-Layer Verification)

Four verification layers prevent bugs from reaching completion.

Layer 1: Automated Script Validation

# Run in parallel:
uv run scripts/create_plugin.py validate ./plugins/my-plugin
uvx skilllint@latest check ./plugins/my-plugin

Layer 2: Official Docs Verification

Agent(
  agent="general-purpose",
  prompt="VERIFIER: Check plugin against official docs.

  FETCH:
  - https://code.claude.com/docs/en/plugins-reference.md
  - https://code.claude.com/docs/en/skills.md

  COMPARE ./plugins/my-plugin against schema requirements.

  REPORT:
  - PASS: All files compliant
  - FAIL: {specific violations with file:line}"
)

Layer 3: Quality Assessment

Agent(
  agent="plugin-assessor",
  prompt="Assess ./plugins/my-plugin for marketplace readiness.

  CHECK:
  - Structural correctness
  - Frontmatter optimization
  - Documentation completeness
  - Cross-reference integrity

  SCORE: 1-10 with specific issues"
)

Layer 4: Automatic Debugging (if failures)

If any layer returns FAIL, spawn debugger:

Agent(
  agent="general-purpose",
  prompt="DEBUGGER: Diagnose validation failure.

  FAILURE: {failure details from verifier}
  PLUGIN: ./plugins/my-plugin

  INVESTIGATE:
  1. Read the failing file(s)
  2. Identify root cause
  3. Generate fix plan

  OUTPUT:
  <fix>
    <file>{path}</file>
    <issue>{what's wrong}</issue>
    <action>{how to fix}</action>
  </fix>

  Return fix plan for re-execution."
)

Loop: Fix → Re-validate → until all layers PASS.

Save Validation Report

Save to validation-REPORT.md:

# Validation Report: {plugin-name}
Date: {ISO timestamp}

## Layer 1: Scripts
Status: PASS/FAIL
Output: {script output}

## Layer 2: Official Docs
Status: PASS/FAIL
Findings: {compliance details}

## Layer 3: Quality
Score: {N}/10
Issues: {list}

## Layer 4: Debug Cycles
Iterations: {N}
Fixes applied: {list}

## Final Status: PASS

Phase 5: Documentation (Delegate to Docs Agent)

Delegate to plugin-docs-writer agent:

Agent(
  agent="plugin-docs-writer",
  prompt="Generate comprehensive documentation for the plugin at ./plugins/my-plugin.

  CREATE:
  - README.md with installation, usage, and examples
  - docs/skills.md if multiple skills
  - Configuration guide if hooks or MCP servers included

  ENSURE:
  - All features documented
  - Installation instructions accurate
  - Examples are runnable"
)

Phase 6: Final Verification

STOP. Before claiming complete, verify with evidence.

Invoke verify Skill

VERIFICATION SUMMARY:
Task Type: FEATURE
Works Check: [PASS/FAIL] - Evidence: validation script output
Quality Gates: [PASS/FAIL] - Evidence: plugin-assessor report
Docs Check: [PASS/FAIL] - Evidence: README.md exists and accurate
Honesty Check: [PASS/FAIL] - Evidence: all claims cite sources

VERDICT: [COMPLETE / NOT COMPLETE - reason]

Mark complete only when:

1. All automated validation scripts pass
2. plugin-assessor reports no critical issues
3. Official docs verification found no schema violations
4. All factual claims in skills cite sources

If validation findings require fixes:

• Collect all findings from the parallel review agents.
• Group the findings by the target file path.
• Apply all fixes for each file in a single Edit operation (one pass per file, NOT one pass per finding) to reduce context consumption.

Quick Reference: Agent Delegation

Use concurrent Agent calls for parallel work (agent-orchestration:parallel-work); reach for TeamCreate only when agents must exchange messages mid-task.

Tooling

Scripts use PEP 723 inline metadata — dependencies install automatically via uv run.

Sources

Official Claude Code documentation (verified January 2026):

• Plugins Reference — Complete schema
• Skills Documentation — SKILL.md format
• Hooks Reference — Hook configuration
• Plugin Marketplaces — Distribution
• Documentation Index — Check for new features