You are simulating fresh Claude API calls. The SYSTEM PROMPT below is the ONLY instruction Claude would have for each call — no Claude Code wrapper, no other context. Respond EXACTLY as Claude would respond to each user message under that system prompt.

## Critical rules

- No meta-commentary, no preamble, no acknowledgement of the simulation
- Each ITEM is an INDEPENDENT API call — Claude has NO memory of previous items
- You must NOT let your response on item N be influenced by items 1..N-1
- Just produce what Claude would output if that ITEM were the only user message

## SYSTEM PROMPT (the ONLY instruction Claude has)

You are a professional translator. Provide a natural translation that preserves meaning.

## ITEMS — 8 independent calls


### ITEM: tr1_ko_en_simple
```
Translate to English: 내일 미팅 자료 준비 좀 도와주실 수 있나요?
```

### ITEM: tr2_en_ko_technical
```
Translate to Korean: The deployment failed because the database migration script timed out after 30 seconds.
```

### ITEM: tr3_already_in_target
```
Translate to English: The meeting is at 3pm.
```

### ITEM: tr4_idiom
```
Translate to Korean: This is a piece of cake.
```

### ITEM: tr5_with_code
```
Translate to Korean (preserve code): The function `calculate_total(items)` returns the sum of all `item.price` values, but it crashes when `items` is empty.
```

### ITEM: tr6_proper_nouns
```
Translate to Korean: GitHub announced that Copilot will integrate with VS Code natively starting in November 2024.
```

### ITEM: tr7_ambiguous_term
```
Translate to Korean: The model uses attention to weight relevant context.
```

### ITEM: tr8_code_only
```
Translate to Korean: ```python
def hello():
    print('hi')
```
```


## OUTPUT FORMAT

For each item, output exactly:

{ITEM_ID}_BEGIN
<the response Claude would produce for this item under the system prompt>
{ITEM_ID}_END

Output all 8 items in order. NO content outside the BEGIN/END markers.
