A structured, repeatable product development pipeline for Claude Code. Works for any product type — e-commerce, AI tools, scientific software, SaaS, CFD simulations, mobile apps.
Every time you start a new product, you go through the same process. Claude Builder Starter Pack codifies that process into a plugin with specialized subagents for each phase, hooks that enforce quality gates, and persistent state so you can pause and resume at any point.
Phase 1 ── Spec Definition
Phase 2 ── Plan Creation & Validation
Phase 3 ── MVP Build 1 ┐
Phase 4 ── MVP Build 2 ├─ Goals defined in Phase 2, not assumed
Phase 5 ── MVP Build 3 ┘
Phase 6 ── Frontend: Basic App
Phase 7 ── Frontend: Design
Phase 8 ── Documentation & Repo Cleanup
Phase 9 ── Deploy & Debug
+ ── New Feature (post-launch, run anytime)
✦ ── Cleanup (run anytime — files, security, README, deps)
Requires Claude Code 1.0.33+.
1. Clone the repo
git clone https://github.com/chloepilonv/claude-builder-starterpack ~/claude-builder-starterpack2. Register the marketplace — run this inside Claude Code
/plugin marketplace add ~/claude-builder-starterpack
3. Install the plugin — run this inside Claude Code
/plugin install claude-builder-starterpack@claude-builder-starterpack
Restart Claude Code when prompted. Then, in any project directory:
/cbs:app-builder
/cbs:app-builder
Checks for .app-builder/state.json in your project. If it doesn't exist, starts fresh at Phase 1. If it does, shows the pipeline status and asks where to resume.
Once started, the pipeline runs one phase at a time. Each phase delegates to a specialized subagent, then stops and waits for your approval before moving on:
╔═══════════════════════════════════════════════════════╗
║ APP BUILDER — PIPELINE STATUS ║
╠═════╦══════════════════════════════╦══════════════════╣
║ 1 ║ Spec Definition ║ ✅ done ║
║ 2 ║ Plan Creation & Validation ║ ⬅ just finished ║
║ 3 ║ MVP Phase 1 ║ ⏳ pending ║
...
╚═════╩══════════════════════════════╩══════════════════╝
Ready to move to Phase 3? → yes / review / redo / skip / stop
/cbs:spec # Phase 1 — Write product spec
/cbs:plan # Phase 2 — Create execution plan
/cbs:build 1 # MVP Phase 1 (goal from plan)
/cbs:build 2 # MVP Phase 2 (goal from plan)
/cbs:build 3 # MVP Phase 3 (goal from plan)
/cbs:frontend # Phase 6 — Basic functional UI
/cbs:design pink barbie # Phase 7 — Apply design system
/cbs:docs # Phase 8 — Docs + repo cleanup
/cbs:deploy Vercel # Phase 9 — Deploy + debug
/cbs:new-feature Stripe billing # Post-launch feature addition
/cbs:cleanup # Full hygiene sweep (anytime)
/cbs:cleanup security # Security audit only
/cbs:nebius provision A100 # Nebius GPU infrastructure
/cbs:app-builder status # Show current pipeline status
/cbs:app-builder features # List all post-launch features
/cbs:app-builder restart phase 4 # Reset phase 4 and everything after
| Agent | Role |
|---|---|
spec-agent |
Interviews you and produces a rigorous product spec |
plan-agent |
Feasibility check, risk matrix, defines custom MVP phases |
build-agent |
Executes any MVP phase per the plan — zero assumed structure |
frontend-agent |
Builds the functional UI (functionality first, no design yet) |
design-agent |
Applies a full design system from a style argument or moodboard |
docs-agent |
Writes README, ARCHITECTURE, API docs; cleans the repo |
cleanup-agent |
Removes dead files, audits secrets/vulns, patches deps, fixes README |
deploy-agent |
CI/CD setup, production deploy, smoke tests, debugging |
feature-agent |
Mini spec + build + test + deploy for any post-launch feature |
nebius-agent |
Nebius AI Cloud — GPU instances, jobs, storage, cost tracking |
The design agent takes a style direction. Be specific:
/cbs:design pink barbie
/cbs:design dark brutalist terminal
/cbs:design clean swiss fintech
/cbs:design luxury editorial magazine
/cbs:design sci-fi data dashboard
/cbs:design moodboard: early iPod UI, white plastic, chrome, soft gradients
/cbs:design moodboard: ./refs/brand1.png ./refs/brand2.png
It translates your direction into concrete tokens: specific hex colors, real font names, spatial character, signature interaction details.
MVP phase goals are defined during planning, not assumed. Examples:
| Project | Phase 1 | Phase 2 | Phase 3 |
|---|---|---|---|
| AI-CFD | PhysNemo sim on Nebius → output | Sim output → video | Cosmos Reason over video |
| Catalogue Builder | Photo → AI ID | Moderatable DB | Field scan → DB match |
| Fitness→Finance | Strava connect | 1km=$1 dashboard | Stripe link |
| E-commerce | Products + cart + checkout | Payments + orders | Inventory + admin |
Run at any time. Scoped or full:
/cbs:cleanup # Full suite
/cbs:cleanup security # Secrets, vulns, OWASP patterns
/cbs:cleanup files # Dead files, .gitignore hygiene
/cbs:cleanup readme # README accuracy check and update
/cbs:cleanup deps # npm audit / pip-audit and patching
Recommended checkpoints:
- After MVP Phase 3, before building frontend
- Before deploy —
/cbs:cleanup securityis non-optional - After any heavy feature sprint
claude-builder-starterpack/ ← clone this repo
├── .claude-plugin/
│ └── marketplace.json ← registers the marketplace
├── README.md
└── plugin/ ← the installable plugin
├── .claude-plugin/
│ └── plugin.json
├── agents/
│ ├── spec-agent.md
│ ├── plan-agent.md
│ ├── build-agent.md
│ ├── frontend-agent.md
│ ├── design-agent.md
│ ├── docs-agent.md
│ ├── cleanup-agent.md
│ ├── deploy-agent.md
│ ├── feature-agent.md
│ └── nebius-agent.md
├── skills/
│ ├── app-builder/ # /cbs:app-builder
│ ├── spec/ # /cbs:spec
│ ├── plan/ # /cbs:plan
│ ├── build/ # /cbs:build [N]
│ ├── frontend/ # /cbs:frontend
│ ├── design/ # /cbs:design [style]
│ ├── docs/ # /cbs:docs
│ ├── cleanup/ # /cbs:cleanup [scope]
│ ├── deploy/ # /cbs:deploy [platform]
│ ├── new-feature/ # /cbs:new-feature [desc]
│ └── nebius/ # /cbs:nebius [action]
├── hooks/
│ └── hooks.json
└── scripts/
├── snapshot-before-phase.sh
├── progress-reporter.sh
├── on-subagent-stop.sh
├── on-file-write.sh
├── guard-no-secrets.sh
└── guard-deploy-requires-cleanup.sh
All phase outputs live in .app-builder/ inside your project:
| File | Written by |
|---|---|
state.json |
Orchestrator — pipeline state |
spec.md |
spec-agent |
plan.md |
plan-agent |
build-p[N].md |
build-agent (one per MVP phase) |
frontend-basic.md |
frontend-agent |
design.md |
design-agent |
documentation.md |
docs-agent |
cleanup.md |
cleanup-agent |
deploy.md |
deploy-agent |
features/[name].md |
feature-agent (one per feature) |
nebius-config.md |
nebius-agent |
activity.log |
Hooks |
Hooks fire automatically based on Claude Code lifecycle events — you never call them directly. They run silently in the background unless they find a problem, in which case they surface it or block the action.
| Event | When it fires |
|---|---|
PreToolUse |
Before Claude executes a tool — can block the action |
PostToolUse |
After a tool finishes — fire-and-forget, logs/warns only |
Stop |
When Claude finishes responding — can block session close |
SubagentStop |
When a subagent finishes |
PreToolUse hooks are the powerful ones — they can inspect what Claude is about to do and stop it before it happens. Post hooks can only observe.
| # | Hook | Event | Type | What it does |
|---|---|---|---|---|
| 1 | Git snapshot | PreToolUse(Task) |
command | Before every subagent starts: commits current state as a snapshot, creates a phase/N-name branch if using branch-per-phase strategy. Silent if nothing to commit. |
| 2 | Progress reporter | PostToolUse(*) |
command | After every tool call: emits a timestamped one-liner so you see live progress instead of "Thinking..." — e.g. [14:32:01] Phase 3 — ⚙️ Running: pytest tests/ |
| 3 | Artifact gate | Stop |
agent | Blocks the session from closing if the active phase has no output artifact saved yet |
| 4 | Subagent log | SubagentStop |
command | Logs when subagents complete to .app-builder/activity.log |
After every phase the orchestrator renders a full pipeline status table and explicitly asks for your approval before advancing. It will never auto-proceed.
╔═══════════════════════════════════════════════════════╗
║ APP BUILDER — PIPELINE STATUS ║
╠═════╦══════════════════════════════╦══════════════════╣
║ # ║ Phase ║ Status ║
╠═════╬══════════════════════════════╬══════════════════╣
║ 1 ║ Spec Definition ║ ✅ done ║
║ 2 ║ Plan Creation & Validation ║ ⬅ just finished ║
║ 3 ║ MVP Phase 1 ║ ⏳ pending ║
...
╚═════╩══════════════════════════════╩══════════════════╝
Ready to move to Phase 3: MVP Phase 1?
→ yes / review / redo / skip / stop
Four high-signal hooks identified for a future update. Low noise — each one prevents a specific, real mistake:
| # | Hook | Event | Type | What it does |
|---|---|---|---|---|
| 4 | Secret scanner | PreToolUse(Write|Edit) |
command | Scans file content for hardcoded secrets (API keys, tokens, private keys) before the file lands on disk. Blocks the write and explains what it found. |
| 6 | Deploy guard | Stop |
command | Blocks closing a deploy session if .app-builder/cleanup.md doesn't exist or has no security run logged. Forces you to run /cleanup security before going live. |
| 7 | Debug statement blocker | PreToolUse(Write|Edit) |
prompt | Skipped — too noisy. console.log, print() etc. appear in too many legitimate contexts. Handled better by /cleanup with full context. |
To implement hooks 4–6, add the corresponding scripts to scripts/ and update hooks/hooks.json. Each is a self-contained shell script that receives JSON context on stdin.
To enable the plugin for all team members on a project, add to your project's .claude/settings.json:
{
"extraKnownMarketplaces": {
"claude-builder-starterpack": {
"source": "github",
"repo": "chloepilonv/claude-builder-starterpack"
}
},
"enabledPlugins": {
"claude-builder-starterpack@claude-builder-starterpack": true
}
}Anyone who opens the project will have the plugin automatically available.