ChatGPTに開始前に失敗の仕方を予測させるプロンプト

Understood. I’ll wait for your response before doing the task. Here are the five most likely failure modes, each with a preventive instruction you could add: 1. **Misunderstanding “preserve verbatim”** 、 I might translate placeholders like `[X]` or `{x}` into Japanese if they look like natural text, or alter `--ar`/`85mm` formatting. *Preventive instruction:* “Treat every bracketed, braced, or flag-like token as opaque code; do not translate or reorder them under any circumstance.” 2. **Assuming you meant “naturalize” Japanese** 、 I might rephrase quoted in-image text to sound more idiomatic, breaking the literal string you want kept. *Preventive instruction:* “For any text inside quotes, output the exact Japanese equivalent of the original wording, word-for-word, even if it sounds stiff.” 3. **Getting generic on structure** 、 I might merge or split paragraphs, or convert bullet points into prose, losing the 1:1 line mapping. *Preventive instruction:* “Match every original line, sentence, and paragraph break exactly; do not combine or split any unit.” 4. **Hedging on technical terms** 、 I might choose a softer or more explanatory Japanese term for things like `24fps` or `4K` instead of keeping the standard loanwords. *Preventive instruction:* “Use standard Japanese technical loanwords (e.g., フレーム/秒, 4K) exactly as they appear in professional contexts; do not substitute synonyms.” 5. **Genuinely hard: handling mixed punctuation and diacritics** 、 I might accidentally convert an em-dash into a comma (you asked for `、`), or drop a diacritic on a foreign name, or misplace a colon inside a JSON value. *Preventive instruction:* “After translating, run a self-check: verify every em-dash is replaced with `、`, every diacritic is intact, and every JSON key/value pair remains structurally identical to the source.” Now I’ll wait for your go-ahead.