Create Plugin
Scaffold a new Claude Code plugin from scratch.
When to use
When you want to create a new plugin that extends Claude Code with skills, commands, and agents. This generates the correct directory structure and wires up MCP tools.
Steps
- Get plugin name and description from the user
- Check for conflicts — call
mcp__plugin_ruflo-core_ruflo__transfer_plugin-search to ensure the name isn't taken
- Create directory structure (follows the canonical plugin contract from sibling plugins' ADR-0001s):
plugins/<name>/
├── .claude-plugin/
│ └── plugin.json
├── skills/
│ └── <skill-name>/
│ └── SKILL.md
├── commands/
│ └── <command-name>.md
├── agents/
│ └── <agent-name>.md
├── docs/
│ └── adrs/
│ └── 0001-<name>-contract.md # Plugin-level ADR (Proposed)
├── scripts/
│ └── smoke.sh # Structural contract (≥8 checks)
└── README.md # Compatibility + Namespace coordination + Verification + ADR sections
- Generate plugin.json with name, description, version, repository (the repo URL the plugin ships from), author (do NOT include
skills, commands, or agents arrays — Claude Code auto-discovers these from directory structure)
- Generate SKILL.md files with proper frontmatter:
---
name: skill-name
description: What this skill does
allowed-tools: mcp__plugin_ruflo-core_ruflo__tool1 mcp__plugin_ruflo-core_ruflo__tool2 Bash
---
- Generate command files with name and description frontmatter
- Generate agent files with name, description, and
model: sonnet
- Generate README.md with install instructions, features, commands, skills, AND the canonical plugin-contract sections:
- Compatibility — pin to
@claude-flow/cli v3.6 major+minor
- Namespace coordination — claim a kebab-case
<plugin-stem>-<intent> namespace; defer to ruflo-agentdb ADR-0001 §"Namespace convention"
- Verification —
bash plugins/<name>/scripts/smoke.sh
- Architecture Decisions — link to ADR-0001
- Generate ADR-0001 (Proposed) at
docs/adrs/0001-<name>-contract.md documenting: pinning, namespace coordination, MCP-tool surface count if applicable, smoke contract scope. Status: Proposed.
- Generate scripts/smoke.sh — at minimum 8 structural checks: version + keywords; skills/agents/commands present with valid frontmatter; v3.6 pin in README; namespace coordination block in README; ADR exists with status
Proposed; no wildcard tools in skills.
- Update marketplace.json if adding to the ruflo marketplace.
MCP-tool drift to avoid (per sibling-ADR lessons learned)
Several plugins shipped with subtle MCP bugs the loop has been finding. Don't replicate them:
embeddings_embed does not exist. Real tool is embeddings_generate. Don't reference embeddings_embed in any allowed-tools line.
agentdb_hierarchical-* does NOT route by namespace. It routes by tier (working|episodic|semantic). Pass tier, not namespace. For namespaced reads/writes, use memory_* instead.
agentdb_pattern-* does NOT route by namespace. It routes through ReasoningBank. Don't pass a namespace arg — fallback writes to the reserved pattern namespace via memory-store-fallback.
pattern (singular) and patterns (plural) are different namespaces. ReasoningBank fallback writes to pattern; hooks_pretrain writes to patterns. Don't conflate them.
Plugin.json schema
Required fields (the repo gate scripts/audit-plugin-manifest.mjs fails without them):
name — plugin identifier (kebab-case, must equal the directory name)
description — what the plugin does
version — semver (X.Y.Z, optional pre-release suffix); bump it on every change, because marketplace users only receive a plugin when its version changes
keywords — non-empty array of strings
repository — the repo URL the plugin ships from (https://github.com/ruvnet/ruflo for plugins in this repo), as a string or an object's url
Recommended fields:
author — { "name": "...", "url": "..." }
homepage, license
Optional fields:
Do NOT include skills, commands, or agents arrays in plugin.json — these are auto-discovered from the directory structure by Claude Code and will cause validation errors if present.
Available MCP tools to wire
Browse available tools: mcp__plugin_ruflo-core_ruflo__transfer_plugin-info
Common tool categories:
memory_* — storage, search, retrieval
agentdb_* — 20 controller-bridge tools (the count grows) (do NOT pass namespace arg — they route by tier or ReasoningBank); call agentdb_controllers at runtime for the canonical list
neural_* — neural training and prediction
hooks_* — lifecycle hooks and intelligence
browser_* — browser automation
workflow_* — workflow management
aidefence_* — safety scanning
embeddings_* — 10 vector-embedding tools (use embeddings_generate, NOT embeddings_embed which does not exist)
1---2name: create-plugin-23description: Scaffold a new Claude Code plugin with proper directory structure, plugin.json, skills, commands, and agents4argument-hint: "<plugin-name>"5allowed-tools: mcp__plugin_ruflo-core_ruflo__transfer_plugin-info mcp__plugin_ruflo-core_ruflo__transfer_plugin-search mcp__plugin_ruflo-core_ruflo__transfer_store-search Bash Read Write Edit6metadata:7 source: https://skillmd.com/skills/ruvnet/create-plugin-28---910# Create Plugin1112Scaffold a new Claude Code plugin from scratch.1314## When to use1516When you want to create a new plugin that extends Claude Code with skills, commands, and agents. This generates the correct directory structure and wires up MCP tools.1718## Steps19201. **Get plugin name and description** from the user212. **Check for conflicts** — call `mcp__plugin_ruflo-core_ruflo__transfer_plugin-search` to ensure the name isn't taken223. **Create directory structure** (follows the canonical plugin contract from sibling plugins' ADR-0001s):23 ```24 plugins/<name>/25 ├── .claude-plugin/26 │ └── plugin.json27 ├── skills/28 │ └── <skill-name>/29 │ └── SKILL.md30 ├── commands/31 │ └── <command-name>.md32 ├── agents/33 │ └── <agent-name>.md34 ├── docs/35 │ └── adrs/36 │ └── 0001-<name>-contract.md # Plugin-level ADR (Proposed)37 ├── scripts/38 │ └── smoke.sh # Structural contract (≥8 checks)39 └── README.md # Compatibility + Namespace coordination + Verification + ADR sections40 ```414. **Generate plugin.json** with name, description, version, repository (the repo URL the plugin ships from), author (do NOT include `skills`, `commands`, or `agents` arrays — Claude Code auto-discovers these from directory structure)425. **Generate SKILL.md files** with proper frontmatter:43 ```yaml44 ---45 name: skill-name46 description: What this skill does47 allowed-tools: mcp__plugin_ruflo-core_ruflo__tool1 mcp__plugin_ruflo-core_ruflo__tool2 Bash48 ---49 ```506. **Generate command files** with name and description frontmatter517. **Generate agent files** with name, description, and `model: sonnet`528. **Generate README.md** with install instructions, features, commands, skills, AND the canonical plugin-contract sections:53 - **Compatibility** — pin to `@claude-flow/cli` v3.6 major+minor54 - **Namespace coordination** — claim a kebab-case `<plugin-stem>-<intent>` namespace; defer to ruflo-agentdb ADR-0001 §"Namespace convention"55 - **Verification** — `bash plugins/<name>/scripts/smoke.sh`56 - **Architecture Decisions** — link to ADR-0001579. **Generate ADR-0001 (Proposed)** at `docs/adrs/0001-<name>-contract.md` documenting: pinning, namespace coordination, MCP-tool surface count if applicable, smoke contract scope. Status: `Proposed`.5810. **Generate scripts/smoke.sh** — at minimum 8 structural checks: version + keywords; skills/agents/commands present with valid frontmatter; v3.6 pin in README; namespace coordination block in README; ADR exists with status `Proposed`; no wildcard tools in skills.5911. **Update marketplace.json** if adding to the ruflo marketplace.6061## MCP-tool drift to avoid (per sibling-ADR lessons learned)6263Several plugins shipped with subtle MCP bugs the loop has been finding. Don't replicate them:6465- **`embeddings_embed` does not exist.** Real tool is `embeddings_generate`. Don't reference `embeddings_embed` in any `allowed-tools` line.66- **`agentdb_hierarchical-*` does NOT route by namespace.** It routes by tier (`working|episodic|semantic`). Pass `tier`, not `namespace`. For namespaced reads/writes, use `memory_*` instead.67- **`agentdb_pattern-*` does NOT route by namespace.** It routes through ReasoningBank. Don't pass a `namespace` arg — fallback writes to the reserved `pattern` namespace via `memory-store-fallback`.68- **`pattern` (singular) and `patterns` (plural) are different namespaces.** ReasoningBank fallback writes to `pattern`; `hooks_pretrain` writes to `patterns`. Don't conflate them.6970## Plugin.json schema7172Required fields (the repo gate `scripts/audit-plugin-manifest.mjs` fails without them):73- `name` — plugin identifier (kebab-case, must equal the directory name)74- `description` — what the plugin does75- `version` — semver (`X.Y.Z`, optional pre-release suffix); bump it on every change, because marketplace users only receive a plugin when its version changes76- `keywords` — non-empty array of strings77- `repository` — the repo URL the plugin ships from (`https://github.com/ruvnet/ruflo` for plugins in this repo), as a string or an object's `url`7879Recommended fields:80- `author` — `{ "name": "...", "url": "..." }`81- `homepage`, `license`8283Optional fields:84- `graph_adapter` — ADR-130 graph intelligence contract (commented out by default in generated output):85 ```json86 // "graph_adapter": {87 // "edgeRelations": ["my-relation-type"],88 // "nodeTypes": ["entity"],89 // "autoRegister": true90 // }91 ```92 When `autoRegister: true`, the plugin's edges are automatically included in `graph_edges` writes93 by the core graph layer. Declare `edgeRelations` — the relation types this plugin produces.9495**Do NOT include** `skills`, `commands`, or `agents` arrays in plugin.json — these are auto-discovered from the directory structure by Claude Code and will cause validation errors if present.9697## Available MCP tools to wire9899Browse available tools: `mcp__plugin_ruflo-core_ruflo__transfer_plugin-info`100101Common tool categories:102- `memory_*` — storage, search, retrieval103- `agentdb_*` — 20 controller-bridge tools (the count grows) (do NOT pass `namespace` arg — they route by tier or ReasoningBank); call `agentdb_controllers` at runtime for the canonical list104- `neural_*` — neural training and prediction105- `hooks_*` — lifecycle hooks and intelligence106- `browser_*` — browser automation107- `workflow_*` — workflow management108- `aidefence_*` — safety scanning109- `embeddings_*` — 10 vector-embedding tools (use `embeddings_generate`, NOT `embeddings_embed` which does not exist)