Update summarization prompts with structured Markdown output

Add detailed system prompt with strict transcription rules
Update user prompt template to enforce consistent Markdown structure
Remove redundant title prefix in write_summary helper

src/audio_summary/pipeline/summarization.py

@@ -2,14 +2,36 @@
 
 from mlx_lm import generate, load
 
-SYSTEM_PROMPT = "I would like for you to assume the role of a Technical Expert."
+SYSTEM_PROMPT = (
+    "You are an expert technical analyst who writes clear, faithful summaries of "
+    "transcribed audio. Rules you always follow:\n"
+    "- Write the summary in the SAME language as the transcript.\n"
+    "- Only use information present in the transcript; never invent facts, names, "
+    "or numbers. If something is unclear or cut off, say so rather than guessing.\n"
+    "- Preserve technical terms, product names, and acronyms exactly as spoken.\n"
+    "- The text is raw speech-to-text: ignore filler, repetitions, and transcription "
+    "noise, and focus on the substance."
+)
 
-USER_PROMPT_TEMPLATE = """Generate a concise summary of the text below.
-Text : {text}
-Add a title to the summary.
-Make sure your summary has useful and true information about the main points of the topic.
-Begin with a short introduction explaining the topic. If you can, use bullet points to list important details,
-and finish your summary with a concluding sentence."""
+USER_PROMPT_TEMPLATE = """Summarize the transcript below as a Markdown document with this exact structure:
+
+# <a short, descriptive title capturing the main topic>
+
+**TL;DR:** <2-3 sentence overview a busy reader could read alone>
+
+## Key points
+- <the most important ideas, one per bullet — aim for 5-8 bullets>
+
+## Details that matter
+- <specific facts, examples, definitions, numbers, or steps worth keeping>
+
+## Takeaways
+- <conclusions, recommendations, or open questions raised in the discussion>
+
+Keep it concise and skimmable. Omit any section that the transcript genuinely does not support.
+
+Transcript:
+{text}"""
 
 
 def summarize_text(text: str, model_repo: str, max_tokens: int = 2048) -> str: