claude-workflow-v2: A Universal Claude Code Plugin with Agents, Skills, Hooks, and Commands

What claude-workflow-v2 Adds to Claude Code

Claude Code ships without specialized subagents, domain-specific knowledge modules, or automation hooks. claude-workflow-v2 fills that gap with a structured package of four component types. Agents are specialized subagents that activate based on the task context: a debugger for broken functionality, a code-reviewer for staged changes, an orchestrator for large multi-step features, and others. Commands are slash commands the user invokes manually. Skills are on-demand knowledge domains that load automatically when the agent detects the relevant task type. Hooks are scripts that run automatically on Claude Code lifecycle events such as pre-edit, post-edit, and task completion.

The README describes the component counts: 7 agents, 26 commands, 14 skills, and 14 hooks. All four types are bundled in the .claude-plugin/ directory at the repository root.

The plugin supports the skills.sh ecosystem, which means the skills component installs into Claude Code, Cursor, Codex, Windsurf, Cline, and the other 35+ agents that skills.sh supports. The full plugin install (agents, commands, skills, and hooks) is Claude Code-specific.

How Agents Activate Automatically

Agents in claude-workflow-v2 are not invoked by name; they activate based on what the user asks. The README shows three examples: a request about a broken login activates the debugger agent, a request to review changes activates the code-reviewer agent, and a request to add authentication activates the orchestrator agent, which then spawns additional subagents including an architect and a test-writer.

This pattern means the user does not need to remember which agent handles which task type. The underlying mechanism is Claude Code's plugin agent routing, which matches the user's intent against the agents defined in the plugin.

The orchestrator agent illustrates the multi-agent pattern: it breaks down a feature request into subtasks, assigns each to a specialized subagent (architect, test-architect, docs-writer), and coordinates the results. The README's example shows it decomposing an authentication feature into schema design, JWT middleware implementation, endpoint creation, test writing, and documentation update.

The code-reviewer agent demonstrates the verification pattern: it analyzes changed files, reports correct logic, and flags specific issues like missing null checks or missing rate limiting on a given endpoint.

Installation: Five Options from npx to the Agent SDK

The README documents five installation methods. The recommended path for any agent via skills.sh is one command:

npx skills add CloudAI-X/claude-workflow-v2

This installs the skills component into Claude Code, Cursor, Codex, Windsurf, Cline, and the other 35+ supported agents. For the full Claude Code plugin (agents, commands, skills, and hooks), the npx installer is the direct path:

npx install-claude-workflow-v2@latest

For per-session use or development, the repository can be cloned and loaded with the --plugin-dir flag:

git clone https://github.com/CloudAI-X/claude-workflow-v2.git
claude --plugin-dir ./claude-workflow-v2

For programmatic use via the Agent SDK, the plugin can be loaded as a local plugin:

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Hello",
  options: {
    plugins: [{ type: "local", path: "./claude-workflow-v2" }],
  },
})) {
}

A fifth option, the plugin marketplace, is described in the README but listed as available when available, indicating it is not yet fully operational. After loading the plugin, the README instructs verifying via the /plugin command, navigating to the Installed tab to confirm project-starter is listed, and checking the Errors tab for any load problems.

Commands: Git Workflow, Verification, and Output Modes

All 26 commands use the /project-starter: prefix. Four commands set the output mode for Claude's responses: /project-starter:architect for architecture-first analysis, /project-starter:rapid for fast iteration, /project-starter:mentor for explanation-focused responses, and /project-starter:review for strict code quality mode.

The git workflow commands cover the inner loop: /project-starter:commit generates a conventional commit message from staged changes, /project-starter:commit-push-pr chains commit, push, and pull request creation, and /project-starter:sync-branch rebases or merges from main.

The verification commands spawn multiple subagents in parallel: /project-starter:verify-changes runs a build-validator, a test-runner, a lint-checker, and a security-scanner concurrently. The README shows the expected output format: each subagent reports pass, fail, or warning with a specific finding.

The README also lists onboarding and documentation commands: /project-starter:explain-codebase, /project-starter:find-tech-debt, and /project-starter:generate-docs, among others.

Hooks: Automation on Claude Code Events

The 14 hooks run automatically without any user action. Three hook categories are documented in the README: security hooks that run pre-edit to block secrets before they are written to files, formatting hooks that run post-edit to apply prettier or black to modified files, and notification hooks that send desktop notifications when Claude is waiting for input or when a task completes.

The security block is the most consequential: if a pre-edit hook detects a pattern matching an API key (such as sk-... ), it blocks the edit and prints a message identifying the file, line, and pattern. The README shows this as a hard block rather than a warning.

Auto-formatting runs after every edit: the README shows prettier being applied to .tsx files and black to .py files automatically. This removes a common friction point in AI-assisted editing where the agent writes syntactically correct but unformatted code.

Desktop notifications address the context-switching cost of waiting for Claude: the notification tells the user when input is required and when a task finishes, allowing them to do other work while Claude runs.

Skills System and Limitations

Skills provide domain knowledge that loads automatically when the task context matches. The README shows two examples: the designing-architecture skill activates when the user asks about service structure and recommends hexagonal architecture with payment providers as adapters; the optimizing-performance skill activates for performance questions and suggests database indexes, response caching, and pagination.

The skills component is the part of the plugin that works across all 35+ agents via skills.sh. The agents, commands, and hooks require Claude Code specifically.

The primary limitation is that all 26 commands use the project-starter: prefix, which reflects the plugin's internal name in package.json (name: project-starter). The README does not document renaming the prefix for a different project identity, so teams who want branded commands would need to fork and modify the plugin.

The plugin marketplace installation path is described as planned but not yet available as of the last push on 2026-08-25. Teams who want to distribute this plugin internally cannot yet use the marketplace route and would need the npx or clone approach.

The repository is MIT-licensed with a CONTRIBUTING.md, CODE_OF_CONDUCT.md, and TERMS.md indicating external contributions are welcome. A PRIVACY.md and SECURITY.md are also present. The repository root contains a CODEBASE.md documenting the plugin's own structure, useful for contributors navigating the agents/, commands/, hooks/, and skills/ directories, and a PERMISSIONS.md documenting which Claude Code permissions the plugin requests at runtime.

Editorial conclusion

claude-workflow-v2 is worth installing for engineers who use Claude Code or another supported AI agent on a daily basis and want a pre-built set of commands, security hooks, and code-review subagents without building them from scratch. The plugin is MIT-licensed and works across over 35 agents via skills.sh. It is less useful for teams with project-specific workflows that differ significantly from the included commands, since the most impactful customization is writing new agents and hooks rather than using the defaults. Verify after installation that no hooks conflict with your project's existing formatter or security scanner configurations.

Frequently asked questions

Official sources

1. CloudAI-X/claude-workflow-v2 on GitHub
2. Issues
3. License: MIT
4. README