Build apps with AI, even if you can't code.
A Claude Code plugin that turns your ideas into working software through a task-driven workflow.
Claude Code compatibility: Follows features up to v2.1.50. See the Claude Code Docs for the latest features.
- Task-driven workflow — plan → implement → ship, one command per stage
- Autonomous agents that research, code, test, and commit for you
- Works with any stack — Bun, npm, Python, Rust, Go, and more
- GitHub-native — issues become task lists, PRs close issues automatically
- Optional Agent Swarm — parallel execution with multiple Claude sessions for complex features
- Claude Code installed
- GitHub CLI (
gh) installed and authenticated - Git configured
| Plugin | Marketplace | What it does |
|---|---|---|
vtsls |
claude-code-lsps |
Go-to-definition, find-references, type lookups for JS/TS |
pyright |
claude-code-lsps |
Same capabilities for Python projects |
rust-analyzer-lsp |
claude-plugins-official |
Go-to-definition, find-references, type lookups for Rust |
frontend-design |
claude-plugins-official |
Generate production-grade UI components with high design quality |
code-review |
claude-plugins-official |
Review pull requests with structured feedback |
code-simplifier |
claude-plugins-official |
CLAUDE.md-aware code simplification (powers /simplify) |
agent-sdk-dev |
claude-plugins-official |
Scaffold and verify Claude Agent SDK applications |
LSP Setup (Important): After installing LSP plugins, add
"ENABLE_LSP_TOOL": "1"to your project's.claude/settings.jsonunderenv. If LSP servers fail to start, clear the plugin cache with/plugin cache clearand restart Claude Code.
If installed from the Claude Code marketplace, you're done. Verify with /plugin → Installed tab.
# Clone
mkdir -p ~/.claude/plugins/marketplaces
cd ~/.claude/plugins/marketplaces
git clone https://github.com/NOGIT007/innovation-basement.git# Add marketplace
/plugin marketplace add ~/.claude/plugins/marketplaces/innovation-basement
# Install
/plugin install coding-plugin@innovation-basement
Restart Claude Code after installation.
The plugin ships default settings that apply automatically:
plansDirectory: "plans"— plans saved to plans/ folderCLAUDE_AUTOCOMPACT_PCT_OVERRIDE: "70"— auto-compact at 70% contextBash(git:*),Bash(gh:*)— git and GitHub CLI permissions- Custom spinner tips for workflow guidance
Add a unique task list ID to your project's .claude/settings.json:
{
"env": {
"CLAUDE_CODE_TASK_LIST_ID": "<your-project-name>-tasks"
}
}/code:setup adds stack-specific permissions (e.g., Bash(bun:*)) and deployment scripts on top of the base settings.
┌──────────────────────────────────────────────────────────────────────────────────────┐
│ Project Setup (once per project): │
│ │
│ /code:bun-init my-app → /code:setup │
│ │ │ │
│ ▼ ▼ │
│ Create Bun + Next.js Detect stack, generate permissions │
│ + Shadcn + Docker + deployment scripts │
│ │
├──────────────────────────────────────────────────────────────────────────────────────┤
│ Feature Development: │
│ │
│ /plan → /code:plan-issue → /clear → /code:implement → /code:finalizer │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ Explore Create new issue Orchestrator runs Merge or PR │
│ the idea OR enrich existing #33 all tasks + cleanup │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ Output: #42 Auto-compact at 70% Close issue │
│ (ctrl+t to view) No manual handover Delete branch │
│ │
├──────────────────────────────────────────────────────────────────────────────────────┤
│ Deployment (after features merged): │
│ │
│ /code:bun-deploy-staging → test → /code:bun-deploy-production yes │
│ │ │ │
│ ▼ ▼ │
│ Deploy to GCP staging Deploy to production │
│ (1 CPU, 512Mi, 0-3 inst) (requires "yes", tests must pass) │
│ │
├──────────────────────────────────────────────────────────────────────────────────────┤
│ Maintenance: /code:cleanup (refactor CLAUDE.md + organize auto-memory) │
│ │
└──────────────────────────────────────────────────────────────────────────────────────┘
# Project setup (once)
/code:bun-init my-saas-app # Bun + Next.js + Shadcn + Docker
/code:setup # Detect stack, generate permissions + scripts
# Feature development
/plan (shift-tab) add dark mode toggle # Explore the idea
/code:plan-issue add dark mode toggle # → Issue #42 created
# OR enrich existing: /code:plan-issue #33
/clear # Free context before implementing
/code:implement #42 # Orchestrator handles everything
/code:finalizer --pr # Create PR for review
# OR: /code:finalizer # Merge directly to main
# Deployment
/code:bun-deploy-staging # Deploy to GCP staging
/code:bun-deploy-production yes # Deploy to production (requires "yes")Interrupted? Just run /code:implement #42 again. Native tasks track progress (ctrl+t to view).
Initialize a new Bun + Next.js + Shadcn/UI project with GCP Cloud Run deployment setup.
/code:bun-init my-appCreates: Bun 1.3.8 + Next.js 14+ + Shadcn/UI + TypeScript + Tailwind + Docker + Bun test runner.
Detect project stack and configure Claude Code settings + deployment scripts.
/code:setupDetects: Bun/npm/pnpm, Python, Rust, Go, Next.js/Vite, Firebase/Docker and generates:
.claude/settings.json— Permissions tailored to your stackscripts/dev.sh— Local developmentscripts/deploy-staging.sh— Staging deploymentscripts/deploy-production.sh— Production (requires "yes" confirmation)
Research codebase, create native tasks, and create or update a GitHub issue.
# Create new issue from description
/code:plan-issue add user authentication
# Enrich existing issue (e.g. from @claude investigation)
/code:plan-issue #33
# Enrich existing issue with extra context
/code:plan-issue #33 focus on the storage layer
# Use spec file as input
/code:plan-issue @SPEC.mdExisting issue mode (#<number>): Fetches issue body + comments, creates native tasks from the findings, and updates the issue in-place with a task breakdown. Bridges GitHub's @claude agent mode with /code:implement.
New issue mode (default): Researches codebase, creates a new GitHub issue with task manifest.
Output: Issue URL + native tasks created (ctrl+t to view)
Launch task execution for all tasks from the issue.
/code:implement #42 # Auto-detect execution mode
/code:implement #42 --team # Force Agent Swarm (experimental)
/code:implement #42 --no-team # Force subagent orchestratorExecution Modes:
| Mode | When | How It Works |
|---|---|---|
| Subagent (default) | < 4 tasks, or many dependencies | Orchestrator spawns implementer per task (proven, lower token cost) |
| Swarm (experimental) | 4+ independent tasks, or --team flag |
Main session leads an Agent Swarm — each teammate is a full Claude session |
Auto-detection picks swarm mode when 4+ tasks exist with 60%+ independence. Override with flags.
Swarm mode requires setup — see Agent Swarm section.
What happens (both modes):
- Validates issue is open
- Creates feature branch
- Runs tasks in parallel (up to 5 concurrent)
- Commits after each task
- Updates GitHub issue status
- Runs
/simplifywhen complete
Resume: Run the same command again. Both modes reconstruct state from TaskList.
Finalize feature: merge or create PR, close issue, cleanup.
/code:finalizer --pr # Create pull request for review
/code:finalizer # Merge directly to main
/code:finalizer --pr 42 # Specify issue numberWhat happens:
- Verifies all tasks complete
- Creates PR (if
--pr) or merges to main - Closes GitHub issue
- Deletes feature branch (local + remote)
Deploy to GCP Cloud Run staging environment.
/code:bun-deploy-stagingConfig: 1 CPU, 512Mi, 0-3 instances, unauthenticated. Requires GCP_PROJECT_STAGING, GCP_REGION, SERVICE_NAME in .env.
Deploy to GCP Cloud Run production. Requires explicit "yes" confirmation.
/code:bun-deploy-production yesSafety: Staging must exist, tests must pass, explicit "yes" required. Config: 2 CPU, 1Gi, 1-10 instances, authenticated, CPU boost.
Generate conventional commit from staged changes.
/code:commitCreate GitHub PR with auto-generated description.
/code:prRefactor CLAUDE.md and organize auto-memory for progressive disclosure. Keeps context files lean and contradiction-free.
/code:cleanupWhat happens:
- Finds all CLAUDE.md files
- Detects contradictions (asks you to resolve)
- Categorizes instructions (Essential, TypeScript, Testing, Git, etc.)
- Flags redundant/stale/obvious items (asks you to confirm)
- Creates
.claude/rules/structure for detailed rules - Writes minimal root CLAUDE.md (target: <50 lines)
- Organizes auto-memory (deduplicates, removes stale entries)
Output: Summary showing before/after line counts and changes made.
Tasks use Claude Code's native task tracking (ctrl+t to view). Each task has a subject, description with file:line references, metadata (issueNumber, verification), status, and blockedBy dependencies.
/code:plan-issue creates tasks from codebase research. /code:implement executes them — subagent mode spawns implementer agents in isolated worktrees, swarm mode uses independent Claude sessions.
pending → in_progress → completed
↘ blocked (needs help)
Verification gates run automatically: SubagentStop hook (subagent mode) and TaskCompleted hook (swarm mode) detect and run your test command. Tasks cannot complete without passing tests.
Subagent Mode (default) Agent Swarm (experimental)
User User
│ │
▼ ▼
/implement #42 /implement #42 --team
│ │
▼ ▼
Task(orchestrator) ← background Main session = swarm lead
│ │
├─┬─ Task(implementer) ← worktree ─┐ ├─ Teammate 1 (claims tasks)
│ ├─ Task(implementer) ← worktree ─┼→ ├─ Teammate 2 (claims tasks)
│ └─ Task(implementer) ← worktree ─┘ └─ Teammate N (max 5)
│ │
├── Task(implementer) ← Task 4 Teammates self-coordinate via
└── Task(simplifier) shared TaskList
│ │
▼ ▼
"Run /finalizer [--pr]" "Run /finalizer [--pr]"
Experimental — multiple independent Claude sessions instead of subagents. Best for complex features with 4+ independent tasks.
Enable in .claude/settings.json:
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}Use flags: --team (force swarm), --no-team (force subagent). Auto-detection picks swarm when 4+ tasks exist with 60%+ independence.
See docs/agent-swarm.md for full setup including tmux config, display modes, keyboard shortcuts, and limitations.
- Keep scope small — One feature at a time
- Trust the process — Tests run automatically, failures get fixed
- Claude Code auto-memory captures learnings automatically — No manual lesson tracking needed
- Use
/code:bun-initfor new projects — Creates full Bun + Next.js + GCP setup - Use
/code:setup— Auto-generates permissions and deployment scripts for your project
| Shortcut | Action |
|---|---|
Ctrl+T |
View task list (see progress during /implement) |
Shift+Tab |
Toggle plan mode (explore before coding) |
Escape |
Interrupt current generation |
Ctrl+C |
Cancel and return to input |
/clear |
Clear conversation context (fresh start) |
/compact |
Manually compact context to free up space |
/cost |
View token usage for the session |
Up Arrow |
Recall previous message |
See docs/ for guides: getting-started, user-guide, git-workflow, desktop, agent-swarm.
MIT — Created by Kennet Kusk