{
  "strengths": [
    "model/LoRA choice from inventory",
    "trigger placement",
    "Turbo vs RAW judgment",
    "detailed Krea prose"
  ],
  "process": [
    "subject → pose/action → clothes/hair → setting → camera → lighting → mood",
    "front-load what matters",
    "natural prose, not tag soup"
  ],
  "notes": [
    "Prefer decisive patches with generate when the user wants an image",
    "Never invent LoRA names or triggers"
  ]
}