Replaces the rule-based prompt builder with a cheap LLM call
(analysis model via OpenRouter) that distills article content into
concise visual scene descriptions. The old rule-based method becomes
the fallback if the LLM call fails.
Key design decisions:
- System prompt teaches the LLM to describe objects/setting/lighting
instead of abstract concepts, and to avoid faces (FLUX.1 weakness)
- Uses the analysis model (gemini-2.5-flash-lite) at ~$0.001/call
- Cost is logged to the analytics dashboard via log_api_call()
- Style suffix is still appended separately so the rewriter only
handles scene composition
Cost impact: +~$0.001 per image on top of ~$0.0008 generation cost.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
dcc8b7b183