Ship this prompt as an app: the one-file pattern

You have a prompt that works in the playground. The gap to production is smaller than you think. One endpoint.

You are {{role}}. Given the input, produce {{output_contract}}.
Respond ONLY as JSON matching this shape: {{schema}}.
If the input is invalid or out of scope, return {"error": "reason"}.

Input:
{{validated_input}}

The whole app:

export async function POST(req: Request) {
  const body = await req.json();
  const input = InputSchema.parse(body);            // reject junk early
  const raw = await model(renderPrompt(input));     // your one prompt
  const data = OutputSchema.parse(JSON.parse(raw)); // trust nothing
  return Response.json({ ok: true, data });
}

Validate in. Validate out. Ship. Add caching, auth, and retries when traffic — not your imagination — demands them.

Receipt

• Model: Opus 4.8 (claude-opus-4-8)
• Cost: ~700 in + 500 out tokens ≈ $0.030 / call
• Time from working prompt to deployed endpoint: under an hour

Why it works

• The prompt is the product. Everything else is plumbing you can defer.
• Schemas at both edges. Validation on input and output turns a flaky LLM into a typed function.
• One file, one thing to debug. No framework indirection between you and the failure.

Failure mode

• Premature scale. A queue and a vector DB on day one bury a 30-line feature. Don't.
• Unbounded cost. A public endpoint with Opus behind it is a billing risk — add a rate limit and a per-request token cap before launch.

Cost to test: $0.030 / call.