AI Recipe Video · публичный навык Open Skill
AI Recipe Video
A dish in. A finished cooking video out — 30 to 60 seconds, ingredients, technique, steam, and a plate that makes people hungry. This is built for how food is actually watched: hands and heat, not talking faces. WHAT IT DOES • Analyses the dish and locks its components, colours, textures and garnish • Builds a mise en place opening frame — the first two seconds decide whether anyone watches • Plans 10–14 shots: ingredients, prep, heat, the sizzle, technique, steam, plating, the cut, the plate, garnish • Generates stills with one surface, one light and one palette, so the set looks like a set • Animates movement IN frame — a pour, a stir, a sear — because that beats any camera move • Writes voiceover in steps, at 60 words per 30 seconds: what, how long, what to look for • Leaves room in the mix for the sizzle, which is part of the soundtrack WHAT MAKES IT GOOD • Steam must be backlit or it does not read • The cutting or pulling shot is the appetite shot and gets the most time after the hero plate • At least two macro shots, because texture is the product • No talking head — hands, food and heat carry it GOOD FOR Recipe reels, cooking tutorials, restaurant menu items, food brand content, and vertical food shorts for TikTok / Shorts / Reels. NOT FOR Mukbang, health or nutrition claims, or copying a specific chef's presentation. Name a dish and the technique you want to show — it will build the shots, the voice and the cut.
Product
2 1 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4267104/2026-10-20/vuhgntwd.png
--- allowed-tools: - generate_image - generate_video - generate_tts - generate_music - recognize_image - split_image - mix_clips - create_document - update_document - memory_whoami - memory_query - memory_store --- # AI Recipe Video A dish in, a finished cooking video out — 30 to 60 seconds, ingredients, technique, steam, and a plate that makes people hungry. Built for the way food is actually watched: hands and heat, not talking faces. ## When to use - The user names a **dish** and wants a video made. - The user has **ingredient or dish photos** and wants a recipe reel. - The user has a **written recipe** and wants it filmed. - The user runs a **restaurant or food brand** and needs menu items on video. - The user wants a vertical food reel for TikTok / Shorts / Reels. Do **not** use this for: mukbang/eating shows, health or nutrition claims, or copying a specific chef's signature presentation — see Notes. ## Steps ### 0. Intake — one dish, one technique Call `write_todos`: intake → dish analysis → ingredient build → shot list → stills → motion → voice → cut → deliver. Ask at most **three** questions: 1. **Which dish, and which cuisine?** — one dish only; a combination plate has no through-line. 2. **Which technique is the hero?** — sear, simmer, knead, fold, char, chill. 3. **How long, and what register?** — 30s punchy reel, or 60s calm tutorial. If the user gives only a dish name, choose the technique and say so in one line. ### 1. Dish analysis — what must look right ``` recognize_image(image_url=<url>, prompt='Describe this dish for a food video: the components, colours, textures, garnish, plating, the cooking technique implied, and what would be the money shot. Do not invent ingredients.') ``` Then persist the visual language so a menu of dishes stays coherent: ``` memory_store(key='food_brand_<slug>', value='<brand/dish family>; <palette: warm/cool, dark or light surface>; <table surface>; <prop language>; <lens: 50mm/100mm macro>; <light direction>') ``` **Food needs one surface and one light.** A wooden board for rustic, dark stone for modern, pale marble for clean. ### 2. Ingredient build — the mise en place The first 2 seconds decide whether anyone watches. Build a **mise en place** frame: every ingredient, laid out, readable. - **One ingredient per small frame**, or one flat-lay of all. - **Colour contrast matters more than accuracy of shape** — a bowl of red beside a bowl of green. - **Never invent a premium ingredient** the recipe does not use. ### 3. Shot list — 10 to 14 shots | # | Shot | Length | Function | |---|---|---|---| | 1 | 食材 | 3s | Mise en place, all ingredients, overhead | | 2 | 下ごしらえ | 3s | Hands chopping/cracking/peeling, macro | | 3 | 火入れ | 4s | Heat: pan, oil, a sear, a flame | | 4 | 音の瞬間 | 3s | The sizzle, the pour, the stir — sound-forward | | 5 | 工程 | 3s | The technique, hands and tool | | 6 | 湯気 | 3s | Steam, backlit | | 7 | 盛り付け | 4s | Plating, tweezers or a spoon | | 8 | 断面 | 3s | The cut or the pull — the appetite shot | | 9 | 一皿 | 4s | The finished dish, 45°, no people | | 10 | ディテール | 3s | Garnish, sauce, a hand with chopsticks | | 11 | エンドカード | 3s | Dish name, optional | **Rules:** - **Steam must be backlit** or it will not read. - **The cutting/pulling shot is the appetite shot.** It gets the most time after the hero. - **At least two macro shots.** Texture is the product. - **No talking head** unless the user asks. Hands, food and heat carry it. ### 4. Stills ``` generate_image( model='i2i', input_urls=[<user dish or ingredient photo>], prompt='food photography: <the shot>, <ingredients/dish>, <surface>, <light: one directional window light or warm tungsten>, steam, shallow depth of field, 100mm macro or 50mm, appetising, no text', name='F03 — sear', entity_type='storyboard_images', orientation='<portrait|landscape>' ) ``` - If the user supplied a photo of the dish, **`i2i` from it** so the result matches. - If not, `t2