Getting Started

Install

There are two paths.

# Path 1 — the version listed on the Anthropic community marketplace
/plugin marketplace add anthropics/claude-plugins-community
/plugin install claude-code-kit@claude-community

# Path 2 — directly from the repository (tracks the latest commit)
/plugin marketplace add This-HW/claude-code-kit
/plugin install claude-code-kit@claude-code-kit

To install the security hooks (protect-sensitive, auto-format, stop-validator) and pre-commit locally, clone the repository and run setup.sh.

git clone https://github.com/This-HW/claude-code-kit
cd claude-code-kit
./setup.sh

Core workflow: from idea to shippable code

The typical flow: brainstorm to surface ambiguity, lock the plan behind a gate, then let development and validation run as an autonomous loop.

brainstorming  ->  /plan-task  ->  /auto-dev  ->  /review  ->  /test
  (diverge)         (lock plan)     (dev pipeline)  (parallel review) (regression check)

1. Brainstorming — surfacing ambiguity before you plan

Before starting a new feature or refactor with loosely defined requirements, brainstorm first. Forcing multiple perspectives (user scenarios, edge cases, trade-offs) into the open surfaces exactly where you’ll need a human gate later.

2. /plan-task — structured task planning

/plan-task add avatar upload to the user profile page

The goal is to remove 100% of requirement ambiguity (the Phase 1 gate). This is where API signatures, file locations, and the definition of done get locked in — the next phase treats this artifact as the implementation contract.

3. /auto-dev — an automated development pipeline

/auto-dev

Takes the planning artifact and drives implementation, integration, and validation automatically. Agents that modify files run in isolated git worktrees to prevent conflicts during parallel work. Large-scale work (10-100+ files) gets chunked with guidance, and anything beyond that is routed to the native ultracode (dynamic workflow), which the user triggers manually.

4. /review — code review (ruff + review-code + security-scan)

/review

Runs linting (ruff), architecture/style review (review-code), and a security scan (security-scan) in parallel — the Phase 3 validation gate. For deeper deliberation, /multi-perspective-review runs a 3-round, 10-persona review.

5. /debug — a 4-phase debug pipeline

Forces reproduce -> isolate -> fix -> verify, preventing the common failure mode of fixing a symptom while leaving the root cause in place.

/debug profile images intermittently break after login

6. /test — run tests and auto-fix failures

/test

Runs tests and automatically fixes failures. Scoped to the files edited in the current session rather than the full suite — full regression runs belong to CI.

What’s next

• Curious about the architecture? Read About.
• For real-world decisions and how the design evolved, see the (Korean) dev log.
• Code and issues live on the GitHub repository.