LaunchKit

Describe an app in a sentence, flip a few toggles, and get a structured Claude Code prompt you can paste straight into a terminal.

• Why
• Features
• Architecture
• Design decisions and trade-offs
• Tech stack
• Quick start
• Configuration
• Project layout
• Tests
• Trademarks
• License

A one-line request like "build me a habit tracker" leaves an agent guessing about quality bars, deployment, licensing and whether to touch the existing repo. Writing the full brief by hand every time is tedious and easy to get wrong.

LaunchKit keeps the boilerplate half of that brief in one place. You supply the idea and the 3 or 4 choices that actually vary; it assembles the rest.

• Name the app and describe it in up to 4000 characters, with a counter that turns amber at the ceiling rather than silently swallowing a long paste. The description gets 60% of the form and 24px text, since it is the only field you write prose into.
• A microphone button inside the description box: tap it and talk instead of typing, which is how a child gets an idea in. Interim words appear as they are heard and settle in place, anything already typed is kept, and the button is not rendered at all in a browser without speech recognition.
• On a phone the app type and feature lists start folded behind their own answer - the selected type's mark, and the marks of whatever is switched on - so the whole form fits on 1 screen. From md up both are open and the toggle is not rendered.
• The output comes in 3 copyable steps, laid out as a chevron rail you click or arrow through, because an agent cannot move itself into a new terminal tab. Step 1 runs where you are: create the repo, add the shell alias derived from the name, open the tab, then stop. Step 2 is the build itself, which you paste into that new tab so its token usage lands on that session.
• Step 3 is the app icon, pasted into the same tab once the build is done. It carries the house image prompt (one glossy 3D subject on a light gradient, no lettering) and says where the finished icon installs for the app type you picked: app/icon.png and the manifest for a web app, the MV3 manifest icons for an extension, the asset catalog for a native app.
• The alias is underscores only, never a dash, since it is used as a shell function name. "Ice Creams" becomes _ice_creams.
• 6 toggles that each change the prompt in a real way: Open Source, Deploy, Public or Private, Audit, Onboard Local App and Auth. 5 of them add their own guidance section. Everything ships on except Auth, since far from every app needs a login.
• 4 of those toggles map to a real Claude Code skill, shown on the switch and emitted as a numbered run order at the end of the prompt: /onboard, /repo-audit, /repo-public-audit and /repo-open-source-audit.
• Pick an app type and the stack comes with it: Web App on Next.js, Chrome Extension on TypeScript and MV3, TUI on Rust, Native on Swift. Each carries its own mark.
• The output pane stays out of the way until you generate, then reveals below the form and scrolls itself into view.
• Read the assembled prompt in a monospace pane and copy it with 1 click.
• Change a setting after generating and the pane flags itself as out of date, so you never copy a prompt that no longer matches the form.
• Keyboard and screen reader friendly: the stack selector is a real radiogroup and the step rail a real tablist, both with arrow-key navigation, every toggle is a labelled switch, and copy results are announced.

3 layers, and the dependency only ever points downward. The page is a server component that renders 1 stateful client component; that component owns the form state and calls 3 pure functions to turn it into text. There is no server code, no database and no network call at runtime, so the whole site prerenders to static files.

The output is split because an agent cannot move itself into a new terminal tab. The name alone drives the setup step; everything drives the build; the name, the idea and the app type drive the icon.

flowchart LR
    N[App name] --> S1["buildSetupPrompt()"]
    N --> S2
    D[Description] --> S2
    T[6 feature toggles] --> S2
    A[App type] --> S2["buildPrompt()"]
    N --> S3
    D --> S3
    A --> S3["buildIconPrompt()"]
    S1 --> P1[Step 1: repo and tab alias]
    S2 --> P2[Step 2: the build]
    S3 --> P3[Step 3: the app icon]
    P1 --> C[Clipboard]
    P2 --> C
    P3 --> C

• Next.js 16 App Router, React 19, TypeScript in strict mode
• Tailwind CSS v4 with shadcn components built on @base-ui/react
• lucide-react icons, official brand marks from simple-icons, Manrope and JetBrains Mono via next/font
• node:test for unit tests, Playwright for the end-to-end flow
• Hosted on Vercel, with GitHub Actions running every check on push

Requires Node 22.6 or newer, since the test runner strips TypeScript types natively.

git clone https://github.com/bunlongheng/launchkit.git
cd launchkit
npm install
npm run dev

Open http://localhost:3046.

Optional. Everything else runs with no configuration at all.

app/
  page.tsx              # server component: header plus the builder
  layout.tsx            # fonts, metadata, Open Graph, manifest
  manifest.ts           # web app manifest for home-screen install
  globals.css           # theme tokens and the page entrance animation
  icon.png              # 512 web icon, also the header mark
  apple-icon.png        # 180 touch icon, flattened for the iOS mask
  opengraph-image.jpg   # 1200x630 share card
components/
  AppBuilder.tsx        # the only stateful component
  CollapsibleStep.tsx   # folds a step behind its answer at phone widths
  DictateButton.tsx     # Web Speech dictation for the description
  BrandIcons.tsx        # generated: official GitHub, Vercel, Next.js, Chrome, Rust, Swift marks
  NameField.tsx         # app name, previewing the tab alias it produces
  DescriptionField.tsx  # textarea plus character counter
  FeatureToggles.tsx    # the 6 switches, each with its mark and skill command
  AppTypeSelector.tsx   # 4 app types, each with its mark and implied stack
  PromptPreview.tsx     # chevron step tabs, the active prompt, copy, staleness pill
  StepLabel.tsx         # numbered section heading
  ui/                   # shadcn primitives
lib/
  buildPrompt.ts        # the entire domain: settings in, prompt out
scripts/
  brand-icons.mjs       # regenerates BrandIcons.tsx from simple-icons
tests/
  buildPrompt.test.ts   # unit coverage of the prompt rules
e2e/
  prompt.spec.ts        # generate, copy, staleness, keyboard

npm run typecheck   # tsc, sources and tests
npm run lint        # eslint
npm test            # node:test unit suite
npm run test:e2e    # Playwright, boots the dev server itself

Point E2E_BASE_URL at a deployed URL to run the same end-to-end suite against it.

The GitHub, Vercel, Next.js, Chrome, Rust and Swift marks are trademarks of their respective owners. They appear here only to label the technology each option targets, and their artwork comes from simple-icons, which releases it under CC0. Their inclusion implies no affiliation or endorsement.

MIT (c) Bunlong Heng