reloadium.com

{
  "type": "object",
  "properties": {
    "slides": {
      "type": "array",
      "description": "1-8 slides, in display order. A single slide is fine for a one-image post.",
      "items": {
        "type": "object",
        "properties": {
          "headline": {
            "type": "string",
            "description": "The slide's main on-screen text — short and punchy."
          },
          "body": {
            "type": "string",
            "description": "Optional smaller supporting text below the headline."
          },
          "label": {
            "type": "string",
            "description": "Optional small uppercase tag above the headline, e.g. 'STEP 1' or 'MYTH'."
          },
          "imagePrompt": {
            "type": "string",
            "description": "The background illustration's subject/scene only — style comes from the shared `style` param, don't repeat it here. Not needed when `imageUrl` is given."
          },
          "imageUrl": {
            "type": "string",
            "description": "Optional: use this photo as the slide's picture instead of generating an illustration — a product image from getShopifyProduct: pass its short `ref` (photo_1, photo_2, …), NOT the long URL (long URLs get corrupted when copied; a wrong character means a 404) — or a `url` from getBestImage, OR an image id (img_N) of one you generated/edited earlier in this conversation (editImage/generateImage; ids from findGeneratedImages). The photo is shown whole (not cropped) with the text beside/below it. Use different photos on different slides, and pick real photographs — skip logos, icons, maps, flags and tiny thumbnails. If you ALSO give imagePrompt, the photo is RE-STAGED by AI instead (so for photos the user wants unedited — 'no AI', 'as is' — give NO imagePrompt): the product is kept as it is and placed in the scene the prompt describes (billed image generation; can subtly alter small details like logos or text). To make real photos 'cooler', PREFER this (imageUrl + imagePrompt on each slide, one call, all slides in parallel) over calling editImage on each photo first — editImage runs one at a time, roughly 30s each. The imagePrompt describes only the SCENE around what the photo already shows — never add a product, tool or part the photo doesn't contain (do NOT ask for a knife in a photo of its pouch: the image model invents a wrong one). Keep the framing of the photo in imagePrompt: for a close-up or detail photo ask for a close-up in a new setting (macro, dramatic lighting), NOT for the whole product — the model invents whatever it can't see. Every re-staged slide is automatically checked against the original photo by a vision model and regenerated (up to 3 attempts) if the product was changed or the scene is poor; the result's `restageChecks` and summary say how each went — report any slide that used the original photo or only partly matched, don't hide it."
          },
          "restageCloseUp": {
            "type": "boolean",
            "description": "Set true ONLY when the user explicitly asks for AI on a close-up / detail photo (a macro, engraving, texture). By default such photos are kept as the real photo even when you give an imagePrompt, because re-staging a detail shot tends to invent or distort the product."
          }
        },
        "required": [
          "headline"
        ]
      }
    },
    "style": {
      "type": "string",
      "description": "Shared artistic style/treatment applied to every slide's background, e.g. 'flat minimalist illustration, warm pastel palette, soft gradients'. Defaults to a plain clean-illustration style if omitted."
    },
    "font": {
      "type": "string",
      "enum": [
        "sans",
        "serif",
        "condensed",
        "handwritten"
      ],
      "description": "Type style for the on-slide text — ALWAYS choose one to fit the brief's tone: 'serif' for luxury / elegant / premium / editorial / wedding / beauty / heritage; 'condensed' (tall uppercase headlines) for sales / sport / fitness / events / bold poster looks; 'handwritten' for casual / personal / friendly / kids / lifestyle; 'sans' only for neutral, technical or corporate topics. The card has a font picker, so the user can also switch it afterwards for free."
    },
    "format": {
      "type": "string",
      "enum": [
        "4:5",
        "9:16",
        "1:1",
        "16:9"
      ],
      "description": "The shape the user asked for (9:16 for Reels/Stories/TikTok/Shorts) — it decides which tab the result opens on; the user can switch tabs afterwards. Defaults to '4:5'."
    }
  },
  "required": [
    "slides"
  ]
}