| name | scaffold-plugin |
| description | Scaffold a new plugin in the toolchest marketplace. Creates plugins/<name>/ with .claude-plugin/plugin.json, README.md, and skills/<skill>/SKILL.md, then patches the root marketplace.json and the root README plugin table. Use when the user asks to add, scaffold, or create a new plugin. |
| argument-hint | <plugin-name> [first-skill-name] |
| disable-model-invocation | true |
| allowed-tools | Read, Write, Edit, Bash, AskUserQuestion |
scaffold-plugin
Add a new plugin to this marketplace, keeping the three sources of truth in sync.
Inputs
$1 — plugin name (kebab-case, required). Becomes the directory name and the name field everywhere.$2 — first skill name (optional, kebab-case). Defaults to the plugin name.
If $1 is missing, ask the user. Do not invent a name.
Before doing anything
Gather these from the user in a single AskUserQuestion call (don't echo back the plugin name — ask only what you don't have):
- One-line description for the plugin. Will appear in
plugin.json, marketplace.json, and the root README row.
- First skill description (one line). Will appear in the first SKILL.md frontmatter and the README row.
Then propose the full file plan in chat before writing — list every file you'll create and every file you'll modify with a one-line summary. Wait for approval.
Plugin shape: skills vs LSP vs agents
This template scaffolds a skill-based plugin. Two other shapes exist; the three-place sync and metadata conventions below apply unchanged to all of them.
- LSP-server plugin (like
ty-lsp): skip skills/<skill>/SKILL.md and the first-skill questions entirely, and instead create .lsp.json at the plugin root (server-name → {command, args, extensionToLanguage}); the root README table's Skill column gets —.
- Agent-bundling plugin (like
writers-room): add agents/<agent>.md files at the plugin root, one per agent — frontmatter needs name and description; tools is optional (note: agents use tools, skills use allowed-tools). Agents are auto-discovered (no plugin.json field) and register as <plugin>:<agent>. If one agent spawns siblings, its Agent(...) allowlist must use the namespaced form — Agent(<plugin>:<agent>, …); bare names break delegation silently. Adding or renaming an agent means updating every sibling allowlist that names it. A skill that needs the user (interview, confirmation gates) stays main-thread — subagents have no AskUserQuestion.
File plan
Create:
plugins/<plugin>/.claude-plugin/plugin.jsonplugins/<plugin>/README.mdplugins/<plugin>/skills/<skill>/SKILL.md
Modify:
.claude-plugin/marketplace.json — append plugin entry to plugins[]README.md — append row to "Available Plugins" table
Templates
Base every field on existing files. The owner string is arimxyer and the license is Apache-2.0 — do not substitute the local git user.
plugins/<plugin>/.claude-plugin/plugin.json
{
"name": "<plugin>",
"description": "<one-line description>",
"version": "1.0.0",
"author": {
"name": "arimxyer"
},
"repository": "https://github.com/arimxyer/toolchest",
"license": "Apache-2.0"
}
plugins/<plugin>/README.md
Minimal — mirror the structure of plugins/sudo-askpass/README.md:
# <Title-Case Plugin Name>
<One-paragraph description of what the plugin does.>
## What It Does
- <bullet>
- <bullet>
## Usage
It can also be invoked directly:
```text
/<plugin>:<skill> <args>
### `plugins/<plugin>/skills/<skill>/SKILL.md`
```markdown
---
name: <skill>
description: <one-line skill description>
---
# <skill>
<Body of the skill — what it does and when to use it.>
Leave optional frontmatter (argument-hint, allowed-tools, metadata) off the starter SKILL.md. Frontmatter style varies across this repo's plugins; the user can add fields when the skill's shape is clear.
If a skill ships supporting files (templates, examples), put them in skills/<skill>/references/ next to that skill — not at the plugin root (see writers-room's MEMO.template.md / BRAND.template.md).
.claude-plugin/marketplace.json
Append to the plugins array, matching the existing object shape exactly:
{
"name": "<plugin>",
"source": "./plugins/<plugin>",
"description": "<one-line description>",
"version": "1.0.0",
"author": {
"name": "arimxyer"
},
"license": "Apache-2.0"
}
Use the Edit tool with a unique anchor (e.g., the closing ] of plugins[] plus the preceding }) to insert. After editing, verify by reading the file back. The prettier hook will reformat it on save — that's expected.
README.md
Append a row to the "Available Plugins" table. Match the existing column order: | <plugin> | /<plugin>:<skill> | <description> |.
Verification
After writing all files, run:
jq -e . .claude-plugin/marketplace.json > /dev/null
jq -e . plugins/<plugin>/.claude-plugin/plugin.json > /dev/null
Both should exit 0. Then list what was created and ask the user to review.
Out of scope
- Don't commit the changes — leave that to the user.
- Don't run
/check-marketplace-sync automatically; offer it as a follow-up if useful.
- Don't fabricate a real implementation for the first skill — the skeleton above is intentional. The user will fill in the body.
