The 287 characters that save you a penny a turn

Show me the prompt

This is the whole system prompt the model gets on every turn in mk0r's VM mode. It lives at src/core/e2b.ts:170 and is exported as DEFAULT_APP_BUILDER_SYSTEM_PROMPT. The repo is open source on GitHub, so you can verify this yourself.

You are an expert app builder inside an E2B sandbox with a Vite + React + TypeScript + Tailwind CSS v4 project at /app. The dev server is running on port 5173 with HMR. You have Playwright MCP for browser testing. Read your CLAUDE.md files for detailed workflow, coding standards, and memory instructions.

287 characters. Roughly 70 tokens by Anthropic's tokenizer. That is the cost of teaching the model what stack it is working with on every single turn.

Why the prompt can be that small

The 287 characters work because the stack does not have to be explained, it has to be present. The E2B sandbox image is built from docker/e2b/e2b.Dockerfile and the relevant block is at lines 77 to 87:

• npm create vite@latest . -- --template react-ts runs at image build time, so package.json, tsconfig.json, src/main.tsx, and index.html all already exist.
• npm install runs in the same build step, so the node_modules tree is in the image. No model output goes to a dependency list.
• Tailwind v4 is installed as a dev dep and wired into vite.config.ts via the @tailwindcss/vite plugin, copied in from docker/e2b/files/app/vite.config.ts.
• A small HMR bridge, src/_mk0rBridge.ts, is layered in so the preview iframe can stream HMR updates to the parent window. The model is told not to touch it.

When a session is claimed, the dev server is already running on port 5173 inside the sandbox. The model's job is to edit src/App.tsx and add files under src/components/. Everything else is the image's problem.

The cost gap, turn by turn

On Claude Haiku 4.5 in May 2026, Anthropic bills $1 per million input tokens, $5 per million output tokens, and $0.10 per million cached input read. Source: platform.claude.com/docs/en/about-claude/pricing. Multiply those rates by the system-prompt sizes and you get a clean per-turn floor on inference cost before the user even prompts.

The first-turn delta on a fresh session is the loudest. v0's ~11,500 token prompt at $1 per million is $0.0115 per first turn. mk0r's ~70 tokens is $0.00007. That is a 165x ratio. After the cache warms, the gap drops to about 16x because both sides bill at the $0.10 per million cached-read rate, but the cache invalidates every time you switch models or the prefix changes. mk0r's prompt does not have a meaningful prefix to invalidate.

The output side, which actually matters more

Output tokens bill at $5 per million on Haiku 4.5, five times the fresh input rate. If a model has to emit package.json, vite.config.ts, tsconfig.json, postcss config, an index.html shell, and a main.tsx before it gets to your actual app, that is a few hundred tokens of boilerplate per project at $5 per million. mk0r's model never writes any of those files. The output is just the edits to App.tsx and new files under src/components/.

There is a second, less obvious savings here. When the model generates boilerplate, it sometimes gets framework details subtly wrong: a missing peer dep, a wrong Tailwind version, a Vite plugin import that does not exist anymore. Each of those mistakes costs a repair turn, and repair turns are full new turns that bill input and output tokens. A predefined stack collapses that entire category of failure because the framework details are validated by the image build, not by the model.

Where this gets honest

A predefined stack is not a free win for every product. Three places the savings story breaks down:

• Stack diversity costs more than tokens. mk0r picked Vite plus React plus TypeScript plus Tailwind v4 plus Playwright. If you want to also support Next.js with server actions, or Svelte, or a Python backend, you either build separate images or you push framework knowledge back into the prompt, and the savings disappear.
• The user has to want the stack you picked. A predefined stack is an opinion. If a user shows up wanting a Remix app, the cheap-prompt story does not help them. mk0r is explicit about this trade in the picker: pick Quick mode for a single-file HTML app, pick VM mode for the React stack.
• Cache hits eat most of the gap. On a hot path with Anthropic prompt caching, even a 12,000 token system prompt only bills $0.0012 per cached turn. That is real but not life-changing. The real win is at first turn, on model switch, and on the rare-but-expensive cache evictions that happen when a session has been idle for a while.

The honest summary is that a predefined stack is a moderate, durable input-side savings, a meaningful output-side savings, and an outsized reliability win. The reliability win is probably the one that pays for itself the fastest, because every avoided repair turn is a ten-to-twenty-x cost win compared to shaving cents on a system prompt.

A back-of-the-napkin per-session number

Assume a typical mk0r session: 20 turns, all on Haiku 4.5, with Anthropic prompt caching. Pretend a competing builder with an ~11,500-token system prompt is identical in every other way. The difference attributable just to the predefined stack:

• Turn 1, cache miss: 11,500 tokens at $1 per million is $0.0115. mk0r: 70 tokens at $1 per million is $0.00007. Delta on turn one: about $0.0114.
• Turns 2 through 20, cache hits: 11,500 tokens at $0.10 per million is $0.00115 per turn, times 19 turns is $0.0218. mk0r's 70 tokens at $0.10 per million is $0.0000133 across 19 turns. Delta: about $0.0218.
• Add a model switch in the middle (cache invalidates): another $0.0115 cache-write penalty. mk0r: zero.

Total system-prompt-only savings on a 20-turn session: roughly $0.04 to $0.05. At one session per minute on a heavy day, that is $60 to $70 of pure input-side savings. The output-side savings, where the model is not emitting boilerplate, are bigger and harder to pin down because they depend on what apps people build. Put together, it is enough to fund the next 10 percent of free anonymous traffic.

Common questions