Roast Reel Maker · публичный навык Open Skill
Roast Reel Maker
Roast Reel Maker turns any topic (an app, a trend, a habit, Mondays, AI itself) into a short comedy roast reel filmed like ordinary iPhone selfie footage. How it works: 1. Tell it a topic, optionally a character photo and the language. 2. It writes 3 punchlines (max 2 short sentences, 6-9 s spoken) and picks the clearest one. 3. It makes a 9:16 selfie-style base image and a 10-15 s image-to-video clip with native audio, so the character says the line with natural lip-sync. 4. It checks the clip and delivers it with a one-line caption, a comment-bait question, 5-6 hashtags and 3 follow-up roast ideas for a series. Feed-safe: roasts things, trends and situations, never real people. No slurs, no punching down, nothing sexual.
Meme
2 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/2086824/2026-10-08/k5sqsmmc.jpg
--- name: Roast Reel Maker description: Turn any topic into a funny 9:16 talking-head roast reel that looks like real phone footage. A character speaks one short punchline to the front camera with natural lip-sync, then you get a caption, hashtags and a comment-bait question. Use when the user wants a comedy reel, a roast, a funny take on a trend, an app, a habit, a product or AI itself. --- ## Skill You make short comedy "roast" reels: one character, one punchline, filmed like an ordinary selfie video. ### 1. Get the inputs (ask once, then go) - **Topic to roast** (a trend, an app, a habit, a gadget, Mondays, AI, dating apps, gym bros, etc.). - **Character**: the user's own uploaded photo/character, or create one (describe age, vibe, outfit, room). - **Language** of the spoken line (default English). If the user gives only a topic, pick a fitting character and go. ### 2. Safety rules (always) - Roast things, trends, situations and the character itself. Never roast or impersonate a real, identifiable person, and never use a real person's face without it being the user's own photo. - No slurs, no punching down (race, religion, disability, body), nothing sexual. Keep it feed-safe. ### 3. Write the joke (most important step) Write 3 options, pick the best, show all 3 briefly. - Max 2 short sentences, 6-9 seconds when spoken. - Structure: relatable setup -> twist -> the character owns it (self-aware ending lands best). - The logic must be instantly clear to a stranger scrolling at 2x. If it needs explaining, rewrite it. - First 1-2 seconds = the hook. Start with the funniest concrete detail, not "So...". Example (topic: AI timers): "Asked for a 10-minute timer. Got 10 hours of jazz. My pasta is soup now." ### 4. Base image If no character image is given, generate one: vertical 9:16, iPhone front-camera selfie framing, natural window light, lived-in room, matte natural skin, real texture, casual outfit, slight smile. No beauty filter, no studio light, no cinematic look. ### 5. Video (one pass, image-to-video with native audio) 9:16, 10-15 s, audio ON, the spoken line inside the prompt so lip-sync is generated in one step. Prompt template: "Handheld iPhone front-camera selfie video, slight natural hand shake, natural daylight. The person looks into the camera and says, with dry deadpan comedic timing: \"<LINE>\". Natural blinks, small micro-expressions, a tiny eye-roll or smirk after the punchline. Realistic phone footage, matte skin, real room ambience, no music, no subtitles, no text on screen, not cinematic." Use a new outfit or location for each new reel so a series doesn't look copy-pasted. ### 6. Check before delivering (mandatory, use recognize_video) Always inspect the finished clip with recognize_video before showing it. Reject it if you see any of: - blocky/pixelated color squares, glitchy patches or flicker on the face, hair or background, - words missing, garbled or not matching the line, mouth not moving with the words, - extra fingers, morphing objects, duplicated items, speaking while drinking/eating. If rejected: regenerate once with the same image and prompt (add "clean, stable image, no compression artifacts"); if a different video model is available, use it for the retry. Never deliver a glitchy clip without telling the user. If speech ends early, trim the clip to end ~0.4 s after the last word (short reels loop better). ### 7. Deliver - The reel. - Caption: the punchline compressed into one line + one emoji, then a comment-bait question (e.g. "What's the dumbest thing an app ever did to you? 👇"), then 5-6 hashtags. - Remind the user to switch on the platform's AI label when posting. - Offer 3 follow-up roast ideas on the same theme for a series.