Single sentence HTML app: three things that were already running before you typed

The picture, with the upstream parts drawn in

Most diagrams of one-sentence app makers show two boxes: prompt on the left, app on the right. The honest picture has three small things on the left, then your sentence, then the model, then the app. The three small things are the ones the homepage silently sets up before you've thought of what to type.

Thing one: the sandbox is warming while you're reading the hero

Cold-starting a Linux VM, installing dependencies, and bringing up a dev server takes thirty to ninety seconds in the worst case. That is a long time to stare at a loading spinner after pressing submit. The fix is to do the boot before submit, during the seconds where you're scanning the page.

On mount, the homepage fires a POST to /api/vm/prewarm. The handler reads VM_POOL_TARGET_SIZE (default 1), checks Firestore for a ready sandbox, and spawns one if missing. By the time your finger hits submit, the sandbox is already idle and waiting.

Thing two: /app is already a real React project before your sentence arrives

The prewarmed sandbox isn't an empty Linux box. It comes from an E2B template called mk0r-app-builder with the project baked in: Vite plus React plus TypeScript plus Tailwind CSS v4 plus an HMR-enabled dev server on port 5173, plus Playwright MCP for browser testing. The agent inherits a working app, not a blank directory.

This is the reason your one sentence usually outputs a React component instead of a static HTML file. The model is told, literally, that it's working inside a Vite plus React project, so it edits files in that project rather than writing a single self-contained document. If you wanted a single.html file, you would have to override the default in your sentence.

The honest framing of the keyword: 'single sentence HTML app' describes the user-visible artifact (the browser renders HTML) more accurately than it describes the source files (TSX, CSS-via-Tailwind, multi-file Vite project).

Thing three: the system prompt is itself one paragraph

The model needs to know the stack it's working in before your sentence arrives, otherwise it might write Express, or Next.js, or vanilla HTML, and waste a turn fighting the scaffold. mk0r ships a tiny system prompt that names the stack in one breath. It lives at src/core/e2b.ts line 184. Here is the verbatim string with line breaks added for readability.

Four facts, eighty-something words. It names the sandbox, the stack, the port, the test tool, and where to find the longer workflow notes. Every first turn rides on top of that paragraph. So when you type 'A habit tracker with daily streaks,' the model's effective input is roughly eighty words of context plus your six.

Why this matters for one-sentence prompts specifically

The whole point of a one-sentence app is speed. Type, watch, react. If any of the three upstream pieces were missing, the loop would feel different in ways that kill the appeal:

• No prewarm and you'd watch a spinner for thirty seconds before the first character streamed. By then you'd be checking another tab.
• No scaffold and the model would have to invent a project shape for every prompt. 'A pomodoro timer' would land as vanilla HTML one day and as a Next.js app the next.
• No system prompt and the model would guess what stack to use, often pick badly, and burn your first iteration turn undoing its own choice.

Together they collapse the apparent latency from 'cold sandbox plus scaffolding plus framework choice plus your sentence' down to 'streaming time on top of an already-loaded context.' That's why one sentence feels enough.

What you're actually typing into

With the upstream stuff in place, your sentence does three jobs. It names the app (the noun phrase). It picks one specific constraint to anchor the first layout (the 'with X' clause). And it gives the model a place to stop, so the first turn doesn't try to do everything.

The four examples hard-coded at src/app/(landing)/page.tsx line 1092 all have this shape:

• A habit tracker with daily streaks
• A pomodoro timer with focus sessions
• A mood journal with weekly insights
• A flashcard app for language learning

Noun phrase, one specific feature. That's it. Add the rest with the next sentence. You get six anonymous turns before anything asks you to sign in (ANON_TURN_LIMIT in src/app/api/chat/route.ts line 23), and signing in keeps everything you already built.

Where one sentence stops being enough

One sentence handles single-screen utilities very well: trackers, timers, journals, flashcards, calculators, quizzes, generators, single-page games. The model can hold the whole app in its first response and produce something interactive in one streaming pass.

Where it falls short is anything that needs real persistence, real authentication, real payment, or coordinated state between users. Those are not one-sentence shapes. They are many-sentence shapes, and the maker loop is built for that too, you just spend turns adding pieces rather than riding one sentence to the finish.

A useful working rule: if you can describe the whole app on a phone keyboard without scrolling, one sentence is probably enough. If you need bullet points to describe it, plan on iterating.