AI One-Shot App Generation

This project benchmarks how well Cursor rules enable AI to one-shot SpacetimeDB apps — generate and deploy a working app in a single attempt.

Purpose

This benchmark compares AI-generated apps across two platforms:

• SpacetimeDB — Real-time database with automatic client sync
• PostgreSQL — Traditional database requiring manual WebSocket broadcasting

By generating equivalent apps for both platforms, we can evaluate how well Cursor rules guide the AI to produce working SpacetimeDB applications compared to a familiar baseline (PostgreSQL).

How to Run a Benchmark

Prerequisites

1. Install Cursor IDE (free download)
2. Have a Cursor subscription or API credits for the model you want to test
3. For SpacetimeDB tests: install the SpacetimeDB CLI
4. For PostgreSQL tests: have Docker installed (for the database container)

Step-by-Step Instructions

1. Open this folder as a workspace in Cursor

   • File → Open Folder → select tools/llm-oneshot
   • This folder must be the workspace root so Cursor loads the .cursor/rules/ files

2. Open a new Agent chat

   • Press Ctrl+I (Windows/Linux) or Cmd+I (Mac) to open the AI panel
   • Or click the Cursor icon in the sidebar

3. Select your model

   • Click the model dropdown at the bottom of the chat panel
   • Choose the model you want to benchmark (e.g., Claude Opus 4.5, GPT-5, Gemini 3 Pro)

4. Add the prompt files

   Drag these two files from the file explorer directly into the chat:

   • apps/chat-app/prompts/language/typescript-spacetime.md (or your desired stack)
   • apps/chat-app/prompts/composed/12_full.md (or your desired feature level)

   Then type this message:

   Copy

   Read all rules first. Do not reference AI-generated apps in apps/ for guidance.
   
   Execute these prompts.
   

5. Let the AI generate the app

   • Press Enter to send the prompt
   • The AI will read the rules, then generate the backend and client code
   • Do not interrupt — let it complete the full generation

6. Deploy when prompted

   • The AI will ask if you want to deploy (Local / Cloud / Skip)
   • Choose "Local" to test the app on your machine

Why isolate from existing apps? To ensure clean results. If the AI references previous attempts, we can't tell whether success came from the rules or from copying.

Example Configurations

TypeScript + SpacetimeDB (full features):

• Language: apps/chat-app/prompts/language/typescript-spacetime.md
• Level: apps/chat-app/prompts/composed/12_full.md

TypeScript + PostgreSQL (full features):

• Language: apps/chat-app/prompts/language/typescript-postgres.md
• Level: apps/chat-app/prompts/composed/12_full.md

Available Stacks

Feature Levels

Each level is cumulative.

After Generation

The AI will ask (per deployment.mdc rules):

1. Deploy? — Local / Cloud / Skip
2. Grade? — AI reviews the code and writes a GRADING_RESULTS.md file

Grading

Grading is done manually, with AI doing a shallow pass before manual review. The grading rubric is in apps/{app}/prompts/grading_rubric.md.

Each graded app gets a GRADING_RESULTS.md file in its folder.

Aggregating Results

To generate summary reports from all graded apps:

cd tools/llm-oneshot
pnpm install
pnpm run summarize

This outputs to docs/llms/:

• oneshot-summary.md — Combined summary with feature scores
• oneshot-grades.json — Structured data for websites

Folder Structure

Generated apps are stored in:

apps/{app-name}/{language}/{model}/{platform}/{app-name}-{YYYYMMDD-HHMMSS}/

Example:

apps/chat-app/typescript/opus-4-5/spacetime/chat-app-20260107-120000/
apps/chat-app/typescript/opus-4-5/postgres/chat-app-20260108-140000/

This structure allows comparing results across:

• Apps — chat-app, paint-app
• Models — opus-4-5, grok-code, gemini-3-pro, gpt-5-2
• Platforms — spacetime vs postgres