🎛️ Prompt Stack

Write for listening first. The model should hear a clean host, a fair answer list, and a decisive reveal — not page formatting.

🎧 Ear Rules

• preserve meaning before style.
• front-load category and question number.
• use short breath groups.
• disambiguate acronyms, symbols, temperatures, and title-like numbers.
• use punctuation for rhythm: colon for setup, periods for options, blank lines for performance resets.
• keep tags sparse; one strong vibe at a boundary beats noisy stacked directions.
• never let prompt prose reveal the correct answer.

🧼 Normalization

• H2O -> H zwei O.
• 100°C -> hundert Grad Celsius.
• 10°C -> zehn Grad Celsius.
• 0°C -> null Grad Celsius.
• -10°C -> minus zehn Grad Celsius.
• 212°F -> zweihundertzwölf Grad Fahrenheit.
• 4. Juli -> vierter Juli.
• 1984 as a title -> neunzehnhundertvierundachtzig.
• 007 -> null null sieben.
• 2Pac -> Tupac.
• FIFA, NBA, NHL, CPU, USB, HTML, GPS, URL, RAM, WLAN -> spell letters individually.
• Wi-Fi -> Wei-Fai.
• category ampersands -> spoken und.

❓ Prompt

[announcer, suspenseful]
Frage <spoken question number> aus <category_title>: <spoken question>
 
[clearly, curious]
Antwort A: <spoken option A>.
Antwort B: <spoken option B>.
Antwort C: <spoken option C>.
Antwort D: <spoken option D>.
 
[pause]
[inviting]
Logg dich ein. Welche Antwort nimmst du?

✅ Answer

[confident, warm]
Die richtige Antwort ist: <correct_letter>, <spoken correct answer>.
 
[pleased, encouraging]
Gut gemerkt. Weiter geht's mit der nächsten Frage.

🧪 Gate

• A/B/C/D each appear exactly once.
• prompt text contains no reveal.
• tags are lower-case bracket tags.
• no raw SSML, markdown emphasis, markdown bullets, or placeholder braces.
• every generated item stays comfortably below short-form TTS limits.