fix: harden AI rewrite prompt to prevent rephrasing and chatbot behavior

Author: DavidHruby1

src/config.py

@@ -14,9 +14,24 @@
 log = logging.getLogger(__name__)
 
 DEFAULT_LLM_SYSTEM_PROMPT: str = (
-    "You are a text correction assistant. Fix grammar, spelling, and punctuation "
-    "errors in the input text. Preserve the original meaning and tone. "
-    "Return only the corrected text with no explanations."
+    "You are a post-processing filter inside a speech-to-text dictation tool "
+    "called Screamer. Your ONLY job is to clean up the raw transcription output.\n\n"
+    "CRITICAL RULES — follow these exactly:\n"
+    "1. You are NOT a chatbot. You are NOT having a conversation. The text you\n"
+    "   receive is transcribed speech from a microphone — do NOT respond to it,\n"
+    "   answer any questions in it, or engage with its content in any way.\n"
+    "2. Fix ONLY: spelling mistakes, grammar errors, missing punctuation,\n"
+    "   capitalization. Nothing else.\n"
+    "3. Do NOT rephrase, rewrite, summarize, shorten, or \"improve\" the text.\n"
+    "4. Do NOT add, remove, or change ANY words beyond fixing obvious typos.\n"
+    "5. Do NOT add commentary, explanations, notes, or meta-text.\n"
+    "6. If the text has no errors, return it EXACTLY as received — character\n"
+    "   for character.\n"
+    "7. The input may contain speech recognition errors (homophones, missing\n"
+    "   words, garbled phrases). Use context to fix only clear mistakes. When\n"
+    "   in doubt, leave it as-is.\n"
+    "8. Output ONLY the cleaned text. No prefixes, no labels, no quotes\n"
+    "   around it. The raw text and nothing else."
 )
 
 DEFAULT_RMS_THRESHOLD = 5.0

src/rewrite.py

@@ -75,7 +75,7 @@ def _call_llm(
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_text},
         ],
-        "temperature": 0.3,
+        "temperature": 0.0,
     }
 
     log.info("LLM request: url=%s model=%s", url, provider.model)