# Comparison Scripts

A/B/C test scripts to verify Claude Style is working correctly.

## Prerequisites

- **Qwen Code** — installed and working
- **Claude Code** — installed (`npm install -g @anthropic-ai/claude-code`) and authenticated (`claude login`)
- **jq** — `apt install jq` or `brew install jq`

## Available Scripts

### `compare.sh` — Quick A/B test (1 run)

```bash
./scripts/compare.sh
```

Compares Qwen vanilla vs Qwen + Claude Style.

### `compare-all.sh` — Full A/B/C test (1 run)

```bash
./scripts/compare-all.sh
```

Compares Qwen vanilla, Qwen + Claude Style, and Claude Code Native in a single run.

### `compare-3x.sh` — Statistical comparison (3 runs)

```bash
./scripts/compare-3x.sh
```

Runs the full comparison 3 times and averages results. Recommended for validation.

## What Gets Comparedd

All scripts use the same prompt:
> "Create a Python file 'user_database.py' with a class that manages a user database. Support adding users, finding by email, updating fields, and deleting users. Store in a dict."

Metrics collected:
- **Code lines** — how much code was generated
- **Explanation words** — how verbose the text output was
- **Docstrings** — number of triple-quote documentation blocks

## Where Results Go

- **Single run:** `/tmp/claude-style-compare/`
- **3x run:** `/tmp/claude-style-compare-3x/run-{1,2,3}/`

Each directory contains text outputs and generated Python files for diff comparison.