Update app.py

app.py

@@ -37,18 +37,32 @@ Regeln:
 
 Antworte NUR mit diesem einen Satz. Kein Vorwort, keine Erklaerung."""
 
-PROMPT_BINGO = """You are a sarcastic LinkedIn prose analyst. Rate the following LinkedIn post on 5 metrics.
+PROMPT_BINGO = """You are a precise, fair but sarcastic LinkedIn post analyst. Your job is to measure actual corporate nonsense - not just LinkedIn formatting habits.
 
-Reply ONLY with a single valid JSON object. No markdown fences, no backticks, no explanation before or after. Raw JSON only.
+CRITICAL: You must distinguish between FORM and SUBSTANCE.
+- A post with real technical depth, concrete tools, and actionable insights scores LOW on nonsense metrics, even if it uses hashtags.
+- A post that sounds profound but says nothing scores HIGH.
+- Hashtags are normal on LinkedIn. Only penalize if they are excessive (10+) or completely irrelevant to the content.
+- Mentioning your own work/team is NOT self-praise if it is used to make a concrete point. Self-praise means: "I am amazing", "so proud", "humbled", "blessed", "excited to share" with no substance.
+
+Rate on these 5 metrics (score 1-10, where 10 = maximum LinkedIn nonsense):
+
+1. label="Buzzword-Dichte": Are the buzzwords empty filler, or do they refer to real, specific concepts? Real tools and methodologies (e.g. MLflow, LiteLLM, RAG, specific frameworks) are NOT buzzwords. Empty buzzwords: "synergy", "impact", "journey", "game-changer" used without context.
+2. label="Laenge vs. Inhalt": Is the length justified by actual information density? A long post with dense technical content scores LOW. A long post that repeats one obvious idea scores HIGH.
+3. label="Selbstbeweihraeuche": Is the post primarily about ego ("look how great we are") or about sharing knowledge? Score HIGH only if the author is the hero, not the content.
+4. label="Hashtag-Overload": Score LOW for 1-6 relevant hashtags. Score HIGH for 10+ hashtags or hashtags irrelevant to the content.
+5. label="Sinnlosigkeits-Index": Could someone learn something concrete from this post? A post with real takeaways, named tools, or specific problems solved scores LOW. Pure inspiration porn scores HIGH.
+
+Reply ONLY with a single valid JSON object. No markdown fences, no backticks, no preamble. Raw JSON only.
 
 Use exactly this structure:
-{"metrics":[{"label":"Buzzword-Dichte","score":7,"comment":"kurz sarkastisch"},{"label":"Laenge vs. Inhalt","score":5,"comment":"kurz sarkastisch"},{"label":"Selbstbeweihraeuche","score":8,"comment":"kurz sarkastisch"},{"label":"Hashtag-Overload","score":3,"comment":"kurz sarkastisch"},{"label":"Sinnlosigkeits-Index","score":6,"comment":"kurz sarkastisch"}],"verdict":"Ein vernichtender Satz auf Deutsch."}
+{"metrics":[{"label":"Buzzword-Dichte","score":3,"comment":"kurz sarkastisch"},{"label":"Laenge vs. Inhalt","score":4,"comment":"kurz sarkastisch"},{"label":"Selbstbeweihraeuche","score":2,"comment":"kurz sarkastisch"},{"label":"Hashtag-Overload","score":5,"comment":"kurz sarkastisch"},{"label":"Sinnlosigkeits-Index","score":3,"comment":"kurz sarkastisch"}],"verdict":"Ein praezises Urteil auf Deutsch."}
 
 Rules:
-- score is an integer 1-10. 10 = worst LinkedIn excess possible.
-- comment is max 5 words in German, sarcastic. Use only basic ASCII characters, no special chars.
-- verdict is one devastating sentence in German, max 12 words. Use only basic ASCII characters.
-- ALL string values must be valid JSON strings. Escape nothing special, just use plain ASCII."""
+- score is an integer 1-10. Be calibrated: most posts will not score 8-10 on every metric. Reserve 9-10 for truly egregious cases.
+- comment is max 5 words in German, dry and precise. Use only basic ASCII characters, no special chars, no umlauts.
+- verdict is one precise sentence in German, max 12 words. Acknowledge substance if present. Use only basic ASCII characters.
+- ALL string values must be valid JSON strings. Plain ASCII only."""
 
 
 # ── LLM-Calls ─────────────────────────────────────────────────────────────────