20k cursed prompts, 266 failures, one LLM rewriter to rule them all

A technical post-mortem of an AI image generation pipeline that ingested over 20k pieces of internet filth, only to choke on 266 of them — mostly Nazi imagery, child exploitation material, and violence. The solution? Throw Claude Haiku at it and hope the rewriter sanitizes the prompts well enough to slip past Gemini's safety filters. It's content moderation as a band-aid strategy: keep the cursed originals, launder the prompts, ship it.