| name | create-plugin |
| plugin | agent-scaffolders |
| description | Scaffolds a new top-level agent plugin directory. NOT for scaffolding single skills (use `create-skill`) and NOT for adding MCP integrations to existing plugins (use `create-mcp-integration`).
|
| argument-hint | [plugin-name] |
| allowed-tools | Bash, Read, Write |
Follow the create-plugin skill workflow to scaffold a new Claude Code plugin.
Inputs
$ARGUMENTS — optional plugin name in kebab-case. Omit to start with discovery.
Steps
- If
$ARGUMENTS provides a plugin name, use it to seed Phase 1
- Follow the create-plugin phased workflow: discover purpose and plugin type,
plan component table (skills / commands / agents / hooks / MCP), ask clarifying
questions per component, scaffold directory structure and
plugin.json (ensuring "author" is an object with "name" and "email"), implement
each component using the appropriate sub-skill, validate, test, and document
- plugin.json Template & Minimal Metadata Standard (MANDATORY — do not skip):
- Ensure
.claude-plugin/plugin.json uses the clean, minimal metadata format:
{
"name": "<plugin-name>",
"version": "0.1.0",
"description": "<description>",
"author": {
"name": "richfrem",
"email": "connect.richfrem@gmail.com"
},
"repository": "https://github.com/richfrem/agent-plugins-skills",
"license": "MIT",
"keywords": [
"<keyword>"
]
}
- Strict Rules:
.claude-plugin/plugin.json must NEVER contain skills, agents, hooks, or commands arrays (these are auto-discovered by Claude Code).
"author" must ALWAYS be an object with "name" (and optionally "email"), never a plain string.
- No duplicate top-level keys.
- plugin.yaml (Hermes compatibility — always generate): After
plugin.json is finalized, scaffold a plugin.yaml at the plugin root for hermes-agent compatibility. Format:
name: <plugin-name>
version: <version>
description: "<description>"
author: <author>
kind: backend
platforms:
- linux
- macos
- windows
provides_tools:
- script_name
skills:
- skill-name
kind: standalone — plugin has no Python scripts that hermes calls directly
kind: backend — plugin has scripts in scripts/ that hermes invokes as tools
- Only include
provides_tools if scripts/ contains callable tool scripts
- Skills list must match actual directory names under
skills/
- Report: "
plugin.yaml created for hermes compatibility. ✅"
__init__.py (Hermes tool/hook wiring — generate when plugin has scripts): If the plugin has callable Python scripts in scripts/, scaffold a root-level __init__.py with a register(ctx) function following this pattern:
from __future__ import annotations
from pathlib import Path
_HERE = Path(__file__).resolve().parent
def register(ctx) -> None:
ctx.register_skill(
name="<skill-name>",
path=_HERE / "skills" / "<skill-name>",
)
- Always include
register_skill() calls for every skill in the plugin
- Only add
register_tool() if the plugin provides callable Python tools
- Only add
register_hook() if the plugin needs lifecycle hooks
- Without
__init__.py, hermes shows "No __init__.py" warning and the plugin won't activate
- Report: "
__init__.py created with register() function. ✅"
- Report the created plugin directory and verification checklist results
Output
Plugin directory with .claude-plugin/plugin.json, component directories, README.md,
and a .claude/settings.json stub for reliable local discovery.
Edge Cases
- If
$ARGUMENTS is empty: begin with Phase 1 discovery — do not pre-fill plugin name
- If similar plugin already exists: reference it as a starting point
- If MCP integrations are needed: invoke
create-mcp-integration for each one
- After scaffolding: run
/agent-scaffolders:audit-plugin to validate structure
Symlink Standards for Shared Scripts
When a skill needs to call a Python helper script that is shared across skills in the same
plugin, always create a file-level symlink in the skill's scripts/ folder pointing to the
canonical copy at the plugin root — never duplicate the file.
Standard pattern:
plugins/<plugin>/scripts/<canonical_name>.py ← canonical source (real file)
plugins/<plugin>/skills/<skill>/scripts/<name>.py ← symlink → ../../../scripts/<canonical_name>.py
The symlink name and target name may differ (e.g. execute.py → exploration_optimizer_execute.py).
The bridge installer resolves all symlinks to physical copies when deploying via the marketplace.
Creating symlinks correctly:
ln -s ../../../scripts/<canonical_name>.py <symlink_name>.py
python plugins/dev-utils/scripts/symlink_manager.py create \
--src plugins/<plugin>/scripts/<canonical_name>.py \
--dst plugins/<plugin>/skills/<skill>/scripts/<symlink_name>.py
⚠️ Windows / core.symlinks warning: If git config core.symlinks is false, git checks
out symlinks as plain-text "stand-in" files. These are silently broken — the bridge installer
copies the path string, not the script. After checkout on Windows or any machine where
symlinks may have degraded, run:
python plugins/dev-utils/scripts/bulk_symlink_fixer.py plugins/<plugin-name>
Then manually verify: find plugins/<plugin-name>/skills -path "*/scripts/*" -type f ! -type l
should return nothing (all script references should be real symlinks, not plain files).
Marketplace Compatibility Note
When this plugin will be distributed via a marketplace.json, the marketplace entry defaults to strict: true, which requires the plugin to have its own plugin.json. A missing plugin.json silently prevents the entire plugin from loading.
Always:
- Scaffold
.claude-plugin/plugin.json inside the plugin directory (this skill does this by default)
- When adding the plugin to a marketplace entry, explicitly set
"strict": true — never rely on the default
- See
manage-marketplace skill for the correct marketplace entry format
References
- Architectural Decision Records (ADRs) located at
references/ADRs/. Always consult them for standards on plugin architecture, shared scripts, cross-plugin dependencies, symlinking, and loose coupling to avoid repeating yourself.