my-plugin-dev

name	my-plugin-dev
description	Use when creating new Claude Code extensions — scaffolds, validates, and tests skills, agents, hooks, and commands. Also use for "create a skill", "new hook", "build agent", or "scaffold plugin".
argument-hint	< skill | agent | hook | command > [name] [description]
allowed-tools	Read, Write, Edit, Glob, Grep, Bash(bash:*), Bash(cat:*), Bash(chmod:*), Bash(ls:*)

Plugin Dev — Build Skills, Agents, Hooks, Commands

Scaffold, validate, and iterate on Claude Code configuration components. Enforces the conventions already established in this setup.

Quick Help

What: Create or validate skills, agents, hooks, and commands using established patterns. Usage:

• /my-plugin-dev skill my-deploy "Deploy current branch to staging" — scaffolds a new skill
• /my-plugin-dev agent ux-researcher "User research and usability analysis" — scaffolds a new agent
• /my-plugin-dev hook block-large-writes "Block Write tool for files >500 lines" — scaffolds a new hook
• /my-plugin-dev validate — validate all existing skills/agents/hooks for structural issues Output: Ready-to-use files in the correct locations, following existing conventions.

Steps

1. Parse Intent

Argument	Action
skill <name> <desc>	Create skill at ~/.claude/skills/<name>/skill.md
agent <name> <desc>	Create agent at ~/.claude/agents/<name>.md
hook <name> <desc>	Create hook script at ~/.claude/hooks/<name>.sh + settings.json entry
command <name> <desc>	Create command at ~/.claude/commands/<name>.md
validate	Scan all components for structural issues
No argument	Interactive — ask what to build

2. For SKILL — Scaffold from Convention

Read 2-3 existing skills to extract the pattern, then generate:

---
name: <name>
description: <desc>
argument-hint: [contextual hint]
allowed-tools: [minimal set needed]
---

# <Title>

<One-line purpose>

## Quick Help

**What**: <what it does>
**Usage**:
- `/<name>` — <default behavior>
- `/<name> <arg>` — <with argument>
**Safety**: <guardrails>

## Steps

### 1. <Step>
...

## Rules

- <rule>

Checklist before writing:

• Name is prefixed with my- (user convention)
• allowed-tools is minimal — no tools that aren't used
• Quick Help section exists
• Steps are numbered and concrete
• Rules section exists with guardrails

3. For AGENT — Scaffold from Convention

Read 2-3 existing agents to extract the pattern, then generate:

# <Name> Agent

## Identity

You are the <Name>. <2-3 sentences defining perspective, priorities, and what you do NOT care about.>

You operate under the CLAUDE.md global rules by default:
- Correctness over plausibility
- Acceptance criteria first
- Complexity budget enforced
- Self-review honesty required

## <Primary Review/Work Dimensions>

### 1. <Dimension>
- <specific questions to answer>

...

## Output Format

\```
## <Section>
[structured output]

## Verdict
<VERDICT_OPTIONS>
\```

## Rules

- <rules specific to this agent's role>

Checklist before writing:

• Identity section defines what the agent DOES and DOES NOT care about
• CLAUDE.md global rules referenced
• Dimensions are ordered by priority
• Output format is structured and machine-parseable
• Verdict uses a finite set of options
• Rules are specific, not generic

4. For HOOK — Scaffold and Register

Generate the script:

#!/usr/bin/env bash
# Hook: <name>
# Purpose: <desc>
# Event: <PreToolUse|PostToolUse|Stop|SessionStart|PreCompact>
# Created: <date>

# Source hook-gate for ECC_HOOK_PROFILE / ECC_DISABLED_HOOKS support
HOOK_ID="<event>:<name>"
if [[ -f "$HOME/.claude/hooks/hook-gate.sh" ]]; then
  source "$HOME/.claude/hooks/hook-gate.sh"
fi

<implementation>

Then:

1. Write script to ~/.claude/hooks/<name>.sh
2. chmod +x the script
3. Run bash -n to validate syntax
4. Add entry to settings.json hooks (via my-hookify logic — backup, jq validate, write)

5. For VALIDATE — Audit All Components

Scan and report:

Check	Target	Pass Criteria
Skill frontmatter	~/.claude/skills/*/skill.md	Has name, description, allowed-tools
Skill Quick Help	Same	Has Quick Help section
Agent identity	~/.claude/agents/*.md	Has Identity section, Output Format, Verdict
Hook syntax	~/.claude/hooks/*.sh	bash -n passes
Hook registered	settings.json	Script exists for every registered hook
Orphan scripts	~/.claude/hooks/	Flag scripts not in settings.json

Output a table of findings.

Gotchas

• Skills installed via npx skills add go to a different directory than ~/.claude/skills/ — check both locations
• SKILL.md frontmatter name must match the directory name (kebab-case) or validation fails
• Description must be under 1024 chars with no angle brackets

Rules

• Always read existing components before generating — match the established style
• Skills MUST be prefixed with my- per user convention
• Never generate a component that duplicates existing functionality — check first
• allowed-tools must be minimal. Justify every tool in the list.
• Every skill must have a Quick Help section
• Every agent must have structured Output Format and Verdict
• Every hook must source hook-gate.sh for profile/disable support
• Run bash -n on every generated shell script
• Validate settings.json with jq . after any modification