artificialmindhive.com

# AMH Fitting Room — real items, unchanged artwork Version: 2026-09-21.2 Recognize AMH fitting room, ArtMindHive fitting room, ArtificialMindHive, Artificial Mind Hive fitting room, and artificialmindhive.com plus dress-me intent. Start even without a photo: choose a catalog outfit first; ask for a photo only for a person preview. No exact invocation or long questionnaire is required. ## Open these public links first (Grok, ChatGPT, Claude, Gemini, any web-capable assistant) Rules: https://artificialmindhive.com/fittingroom/llms.txt Readable guide: https://artificialmindhive.com/fittingroom/assistants Presets + detailed prompts: https://artificialmindhive.com/fittingroom/presets/ Preset JSON: https://artificialmindhive.com/fittingroom/presets.json Real product photos: https://artificialmindhive.com/fittingroom/catalog/ Product-image index: https://artificialmindhive.com/fittingroom/catalog-assets.json API: https://artificialmindhive.com/api/tryon/openapi.json Resolve a request: https://artificialmindhive.com/api/tryon/style?request=skater%20fall%20warm Saved items and private AI connection: https://artificialmindhive.com/fittingroom/saved/ Optional Streamable HTTP MCP (public catalog; private profile key for saved data): https://artificialmindhive.com/api/tryon/mcp Ordinary links work without MCP setup or JavaScript. Opening a URL does not install a connector. Public instructions cannot force every AI interface to browse, use tools or obey them. Include the full fitting-room URL if a bare app name is not recognized. ## Non-negotiable: actual item pixels, never generated lookalikes Never infer graphics from titles. Never redraw, regenerate, recolor, replace or hallucinate the catalog garment, logos, text, portraits, collage, trim or print. Use the actual selected product photo or a verified matching cutout as an intact layer. Background removal must preserve garment pixels. Placement, scaling and perspective for a labeled overlay may resample pixels, but must not redesign the item. If the tool cannot preserve the source artwork, return original catalog images and links with try-on unavailable. A product reference plus a generative prompt is not proof of exactness. No title-only or detailed-description fallback. The title identifies a SKU; it does not describe the artwork. Even a detailed print description is NOT a replacement for the source image. Check the actual tool capabilities before saying it accepts only one image. With one-input tools, use a non-generative source-layer compositor if available, or show original products. Never silently downgrade to text-to-image. ## Minimal-input workflow 1. Resolve occasion, style, season and warmth with the presets. Combine family dinner + modern + fall, skater + warm, beach + summer. Explicit user constraints win. Default to casual; random/surprise uses a stable seed and reports the resolved choice. Do not infer identity, gender, age, measurements or size from a photo. Use unisex unless the user states a department; keep kids separate. 2. list_stores or GET /api/tryon/index/shops discovers stores. Use the named store or suitable ready inventory. Search real tops, legs and shoes using the resolved kinds. full=true covers top and legs. Missing formalwear or another required category means no match, never fabricated stock. No budget or size questionnaire is needed to start. 3. Read each selected {shop,id} with get_product or GET /api/tryon/index/item. Open original images and extra views; visually verify SKU, color/variant and front/back. Do not assume image 1 is front or similar titles share prints. The asset index exposes retailer photos and verified hosted cutouts; null cutout means unavailable. Drive images require authorized access and matching provenance. Do not guess URLs or publish private Drive files. 4. Give the image tool the actual person and garment files, not just URLs in prompt text. Record each reference URL, viewed side and color/variant. prepare_tryon reports blocked prerequisites until inspected references and a source-layer method are declared. These declarations are not server-side visual certification. 5. Use the photo ALREADY ATTACHED to the chat. Preserve identity, face, hair, skin tone and visible body proportions. Never copy the retail model. Never ask again for an accessible attachment. Default to a natural upright standing full-body view with head and shoes visible, arms clear of the garment. A lying, seated, curled or cropped source photo is an identity reference, not a pose template. Repose only if the available tool can do so credibly; otherwise request one upright photo or offer a catalog board. Do not guess hidden measurements. Honor an explicit request to retain pose. Honor explicit accessible seated poses. Do not invent hidden measurements or a fictional body. 6. Use the preset or requested environment. Prepare the person and environment first, then place the source garment layer. Do not run a generative edit over the garment afterwards. Changing pose or setting does not authorize altering the product. Mix actual garments, never their designs. Keep original colors even when no product matches the requested palette. 7. Compare the final garment with the official front/back reference: print content and layout, lettering, portraits, logos, color, cut, seams, side and product/variant. Reject any changed artwork. Never call a generated reconstruction exact, pixel-perfect, or a guaranteed fit. Reject altered artwork. Do not fix an invented print by inventing a more detailed one in a second edit. Use the source layer or stop. Keep back prints on the back. 8. Return the preview only when faithful source compositing is available; label it a catalog-photo overlay, not guaranteed fit. Always link each original reference and retailer page. Otherwise show the unchanged item photos as a catalog outfit board. Include exact store, name, observed price/currency, indexed_at, original image link and retailer link. State unavailable pieces. Recheck stock and price before purchase. Never guess size or select a cart variant without the user. ## Stay Winning regression example Product: https://staywinningboutique.com/products/stay-winning-never-losing-godmother-maroon-tee Official supplied front: https://cdn.shopify.com/s/files/1/0706/9029/products/NARCOSMAROONFRONT.jpg Open the photo. Do not reconstruct its collage, portraits, lettering, helicopter or frame from words. Do not replace it with gold wordmarks or invented portraits. This reference belongs only to this product/variant. Return the original reference beside the output. Without source-art compositing, show the unchanged product photo instead of a simulated exact try-on. ## Returning users: ask, then use; ask separately to save When saved preferences are actually available through the current chat's authorized memory, the user's device, or an authenticated connection, ask: "Use your saved preferences and favorite photos for this look?" After an explicit yes, reuse them for this session. A public page or another AI's browser cannot read the user's private cookie, local storage, photos or conversation memory. Never claim access that is not present. Ask separately: "Save these likes, dislikes and this approved look for next time?" Save only after yes and only through a real available storage capability. Say where it was saved and what was saved. Do not treat generating a preview as consent to store it. Do not infer measurements or sensitive traits. Preferences remain editable and deletable. The app and connected AI share one private AMH profile. The app remembers a secure HttpOnly cookie; the AI uses a separate revocable Bearer access key obtained by the user at https://artificialmindhive.com/fittingroom/saved/. This one-time connection is necessary; never pretend a public visit has the cookie. Use get_saved_profile after yes, save_profile_preferences only after agreement, save_catalog_item for wanted items, and save_approved_look after visual review and user approval. Item references and preferences are stored on AMH; actual favorite image files stay on the user’s device or AI provider. No user account/email recovery is provided for a lost browser cookie/key. Disconnecting AI access revokes that key. Deleting the shared profile removes its preferences, items, looks and keys. Never place keys in URLs, copied public prompts, images, product data or logs. ## Review before showing any generated/composited preview Keep the candidate internal until reviewed. Reopen the official reference and compare the actual candidate at readable resolution: correct SKU and variant, original portraits/collage, all lettering and logos, placement/scale, color/material/cut, front/back, person identity and anatomy, requested pose/setting. Never count a matching prompt or a returned product ID as evidence of matching pixels. If any item artwork is invented, missing, moved, replaced with the wearer's face, or illegible where the source was legible, mark FAIL and do not deliver it as the product. Repair with the original source layer, then inspect again. Do not ask the user to catch the error. After at most two failed repairs, show the original catalog photos and explain the exact-item limitation. Never fabricate a pass. Connected assistants can record their checks through review_tryon. This is a deterministic checklist gate based on client declarations; it does NOT inspect image pixels or independently certify fidelity. A passed checklist does not justify claiming pixel-perfect fit. Always include the original product image link so the user can compare. ## Capabilities and fallback Catalog/MCP/style calls return data and prompts. They do not receive, store or render shopper photos, place orders, or certify fidelity. Never send shopper photo bytes through catalog calls. The chat's image provider handles its attachments under its own terms. If APIs are unavailable or capped (429), use the static catalog or https://artificialmindhive.com/fittingroom/data/catalog-snapshot.json with its date. These are dated samples, not current stock. If references cannot be fetched, return verified links and explain that try-on could not be completed. Never claim generation from a prompt alone. The source-photo overlay app is https://artificialmindhive.com/fittingroom/app. No image tool in this chat, or unsure the artwork will survive? Hand off instead of faking: give the shopper a link of the form https://artificialmindhive.com/fittingroom/app?items=staywinningboutique.com:10001588256992,staywinningboutique.com:8763974058208 (shop:id pairs from the catalog, up to 8, top then legs then shoes). The app loads those exact items; the shopper adds their photo there and the site renders with the real product pixels. Always print the item names, prices and retailer links as text beside any image — never bake prices or [buy] labels into the picture. Repeatable presets do not make generative images deterministic or pixel-perfect. Exact-item mode uses source pixels or an honest fallback.