claude-workflow-plugins

A curated marketplace of production-grade Claude Code workflow skills & plugins by cubha.

Battle-tested skills extracted from a real Claude Code harness — each generalized to run standalone in any project. They span the full loop: plan → research → build → test → debug → review. Korean-first docs (English coming).

Install

# 1. Register this marketplace (catalog only — installs nothing yet)
/plugin marketplace add cubha/claude-workflow-plugins

# 2. Install any plugin à la carte
/plugin install scope@claude-workflow-plugins
/plugin install debugger@claude-workflow-plugins

Adding the marketplace does not install everything — you pick what you want.

Plugins

📐 Planning & Research

Plugin	What it does
scope	Pre-project decision gate: MoSCoW feature scoping + effort sizing + tech-stack confirmation (user-approval gated). Produces a SCOPE-*.md your scaffolding workflow consumes. Fills the gap between "requirements" and "create the project".
team-research	Pre-project research orchestrator: tech-stack / architecture / competitor / UX researcher subagents run in parallel (4 bundled agents), cross-validated into one report.
analyze	Brownfield codebase audit: architecture, tech debt, security, performance → comprehensive MD report with improvement roadmap.

🛠 Development & Debugging

Plugin	What it does
team-dev	Parallel development orchestrator: dependency-analyzes SubTasks, groups them, and implements concurrently in git-worktree isolation (bundled feature-builder agent). Includes an opt-in TDD RED-gate flow.
debugger	Log-driven debugging workflow: insert logs → pinpoint root cause → fix → verify → clean up. --pw mode auto-reproduces web UI bugs via bundled Playwright MCP. JS/TS, Python, Java, Kotlin.
unit-test	Scenario-based functional test orchestrator: git-diff scoping → storyboard → test cases → Playwright MCP execution → PASS/FAIL/WARN report.

🎨 Design Quality

Plugin	What it does
design-lint	Deterministic (no-LLM) design anti-pattern linter — 47 detectors for token hygiene, color/font/spacing consistency, and AI-default clichés (beige bg, text gradients, uppercase eyebrows). Same input → same verdict, 0 tokens → usable as a CI gate (--gate).
init-design	Reverse-extracts a design system from an existing codebase into Ground Truth files (DESIGN-TOKENS.md · UX-BRIEF.md) that downstream UI work can obey.

🧠 Review

Plugin	What it does
braintrust	In-session adversarial review panel: 3–5 independent fork agents re-examine the session's design/plan/decision through different lenses, then reconcile conflicts into a verdict. Context-inheriting red team.

Plugin types

• skill — SKILL.md only: scope, braintrust, analyze, init-design
• skill + bundled agents — team-research (4 researchers), team-dev (feature-builder)
• skill + MCP — debugger, unit-test, design-lint ship a .mcp.json that auto-provisions Playwright MCP (npx @playwright/mcp@latest)

Requirements

• Claude Code (recent version with plugin marketplace support)
• Node.js 18+ for design-lint script and Playwright MCP plugins
• Plugins interoperate but never hard-depend on each other — each runs standalone; cross-references activate only when the counterpart is installed

Repo layout

.claude-plugin/marketplace.json   # catalog (9 plugins)
plugins/<name>/
├── .claude-plugin/plugin.json
├── skills/<name>/SKILL.md        # + references/, scripts/
├── .mcp.json                     # MCP-bundled plugins only
└── agents/*.md                   # agent-set plugins only
verify.sh                         # structural validation (JSON, sources, coupling scan)

License

MIT © cubha — see LICENSE.