thumbnail-generation

# Thumbnail Generation ## Mission Create thumbnail directions that communicate the actual content quickly at small size. Optimize the visual hook and hierarchy, not a misleading promise or an image full of model-generated copy. ## Required inputs - Video/article title or core topic. - Platform and target dimensions; use the platform's current requirements. - Channel/brand style and audience. - Optional source frame, face, product, logo, or other reference image. Collect when relevant: - the actual content promise and approved claims; - subject emotion/pose and what must remain recognizable; - language, overlay copy, safe areas, and number of variants; - budget and whether typography will be added in post-production. ## Model selection Use [MODELS.md](../../MODELS.md) and the live provider contract in [references/muapi-image-tools.md](../../references/muapi-image-tools.md). Use text-to-image for a new concept and image editing when incorporating a real source frame or face. A public MuAPI `youtube-thumbnail` recipe is a useful starting workflow, but its current model and fields must be rechecked. ## Workflow 1. Extract the content hook and write one honest visual promise. Do not invent results, guests, products, or events that do not appear in the content. 2. Propose two or three distinct concepts, such as emotion-first, single-object contrast, before/after only when factual, or a clean curiosity gap. 3. Define the subject, focal point, rule-of-thirds placement, background contrast, safe text region, and no-go elements for each concept. 4. If a local source image is supplied, confirm rights and upload it with `media.upload_file`. Use `media.edit_image` to incorporate or reframe it. 5. Generate one candidate per approved concept in the target aspect ratio. Keep the first batch bounded and state the call count/cost signal first. 6. Poll and preserve each result. Review the thumbnail at small size for immediate subject recognition, contrast, crop, facial/subject drift, artifacting, and truthful content representation. 7. Return suggested overlay copy of three to five words, placement, and post-production typography guidance. Prefer adding exact text outside the image model; if text is generated, verify spelling and legibility. 8. Offer A/B variants only when the user asks or has approved the additional calls. ## Decision rules - One clear focal point beats a crowded collage. - Do not use a fake reaction, badge, celebrity, result, or screenshot to imply content the video/article does not contain. - Keep overlay copy short and high-contrast; text-rendering in generated images is not a substitute for final typography. - Preserve a real source subject rather than regenerating it when identity or product continuity matters. ## Output format Return one row per concept with title/hook, concept label, model/tool, prompt or prompt hash, output URL, status, dimensions, suggested overlay, text placement, QA notes, and receipt link when saved. ## Failure and missing-data behavior If the topic is too abstract to create an honest visual, ask for a more specific angle or source image. If a source or generation branch fails, report it explicitly and do not silently replace it with a generic thumbnail.