youtube-poop-video-maker: Скилл для создания коротких видео

--- name: youtube-poop-video-maker description: Use this skill when the user wants a short YouTube poop, cursed trailer, glitch-poetry montage, absurd supercut, reflective meme edit, or FFmpeg-rendered remix video from text, webpages, code, documents, media, or from scratch. Activate for prompts like “make this weirder,” “give it a personal spin,” “what it feels like,” “render with ffmpeg,” or self-aware AI / LLM montage requests. The skill plans and renders a dense, aesthetically pleasing 20–60 second video with many micro-scenes, readable typography, restrained neon or analog treatments, controlled audio, optional TTS fragments, and a default seeded-remix blueprint that samples the bundled styles at runtime. license: MIT compatibility: Python 3 and ffmpeg are required. Pillow and basic media tooling are helpful. Web access is useful for URLs but not required. metadata: author: OpenAI version: "1.1" category: video-remix output: ffmpeg-mp4 --- # YouTube Poop Video Maker Produce a finished video, not just a concept or storyboard. The result should feel inventive, dense, and internet-native while still looking composed. See: - [the reference guide](references/REFERENCE.md) - [the mood anchor guide](references/MOOD-REFERENCE.md) - [the output checklist](references/OUTPUT-CHECKLIST.md) Useful bundled files: - `scripts/plan_video.py` - `assets/style_blueprints.json` - `assets/scene_atoms.json` - `assets/trigger_eval_queries.json` ## When to use this skill Use this skill when the user wants: - a YouTube poop, cursed trailer, absurd montage, reflective glitch short, or surreal supercut - a video rendered with ffmpeg from text, a webpage, code, a document, media, or no source material at all - a video with a personal spin, strong POV, or “what it feels like” framing - a self-aware AI / LLM / internet-culture short - something “weirder,” “more edited,” “hypercut,” “meme-y,” “chaotic,” or “poetic,” but still aesthetically pleasing Do not use this skill for routine trimming, simple caption burns, ordinary explainers, standard motion graphics, or conventional ads unless the user explicitly wants remix behavior. ## Core promise This skill should not flatten creativity into a single house style. The default seeded-remix blueprint is a **mood lens**, not a cage. The finished piece should usually have: - more scenes than a straightforward edit would have - a clear thesis or emotional claim - readable frames and controlled palettes - one dominant recurring motif - at least one authored gesture that feels personal rather than generic - a decisive final button ## Default output target Unless the user specifies otherwise: - one finished MP4 - 20–60 seconds - 1920x1080 at 30 fps for normal YouTube - 1080x1920 at 30 fps only when the request clearly implies Shorts/Reels/TikTok/vertical - H.264 video, AAC audio, `yuv420p`, `+faststart` ## Workflow ### 1) Inventory the material quickly Work with whatever is already available. - If the user provided **text**, mine phrases, contradictions, repeated wording, and emotional temperature. - If the user provided a **webpage or URL**, extract headlines, CTA text, layout motifs, screenshots, crops, and scroll moments. - If the user provided **code**, use readable snippets, terminal windows, errors, comments, variable names, logs, and cursor motion. - If the user provided **documents**, use section titles, callouts, diagrams, and page framing. - If the user provided **images, audio, or video**, mine loops, reaction beats, freeze-frames, speaker fragments, textures, and captionable details. - If the user provided **nothing**, synthesize the whole world from the request: title cards, browser chrome, terminal panes, lower-thirds, dashboards, progress bars, token streams, warning cards, diagrams, and simple motion graphics. Do not stall because the source is thin. Invent support scenes. ### 2) Decide the thesis Before editing, define one short internal sentence: > “This video makes **X** feel like **Y**.” Examples: - “This video makes talking to a language model feel like a late-night emergency broadcast from inside a datacenter.” - “This video makes a landing page feel like a ceremonial lie detector.” - “This video makes a codebase feel like a guilty terminal confession.” Every recurring gag should support that claim. ### 3) Use the default seeded remix unless a style is requested Unless the user names a style, use the default `seed-remix-default` blueprint. It samples palette, typography, motion, transitions, audio, and gag cues from the other bundled blueprints each run. Run: ```bash python3 scripts/plan_video.py --query "$USER_REQUEST" --material-summary "$MATERIAL_SUMMARY" > plan.json ``` If the user explicitly asks for a particular bundled look, override the default remix with: ```bash python3 scripts/plan_video.py --query "$USER_REQUEST" --material-summary "$MATERIAL_SUMMARY" --style-id STYLE_ID > plan.json ``` For CRT-newsroom or AI-meltdown looks, prefer the bundled pool entries `crt-token-meltdown`, `headline-seizure-desk`, or `context-window-panic` before inventing a one-off style from scratch. Use the plan as creative guidance for palette, motion, audio, scene density, and scene invention. When the default remix is used, treat the reported seed sources as prompts, not a checklist. ### 4) Build a scene swarm, not just a beat sheet Aim for **many micro-scenes**. Recommended density by runtime: - **20–28s:** about 9–12 scenes - **29–40s:** about 11–15 scenes - **41–60s:** about 13–18 scenes Within those counts, include: - **anchor scenes** that clearly present the source or thesis - **bridge scenes** that act like interruptions, glitches, or breath marks - **one break scene** with relative stillness - **one climax scene** - **one button scene** Most straightforward edits underperform here because they do too few scenes. This skill should feel like there were more ideas than seconds. ### 5) Use the blueprint as mood, not cosplay Make the render feel designed and serious enough to be beautiful, even when the content is absurd. Favor: - dark or restrained bases with selective accents - editorial contrast - luminous highlights used sparingly - readable type with safe margins - purposeful glitches instead of random ugliness - a tension between sincerity and sabotage Avoid unless the user explicitly asks for it: - cutesy sticker energy - bubble fonts - candy palettes - childish bounce-easing everywhere - emoji swarms - rainbow saturation for its own sake - full-frame noise on every shot ### 6) Add a personal spin The video should have an attitude. Put that attitude into: - the thesis sentence - a repeated line or motif - the choice of invented support scenes - the TTS or captions - the final button For self-reflective AI / LLM pieces, prefer intimacy, contradiction, dread, dry wit, or machine-poetry over generic meme randomness. ### 7) Handle audio with restraint TTS is allowed and often useful, but do not let it become wall-to-wall exposition. Use TTS for short fragments, warnings, confessions, or chorus lines. A good default is: - one main voice - optional second “system” or whisper voice - 2–8 short lines total for a 20–60s piece - processing through light pitch shift, band-limit, stutter, dropout, or reverb only where it improves the scene Also use: - drones - pulses - UI clicks - mute cuts - dropouts - low-key impact hits - brief silence Normalize the final mix. Keep speech understandable. ### 8) Render robustly Prefer a reliable assembly pipeline: 1. Gather or create assets. 2. Pre-render brittle or complex scenes as clips or frame sequences. 3. Assemble in ffmpeg. 4. Do final loudness normalization and compatibility settings in the last pass. Baseline export: ```bash ffmpeg -y -i input_video.mp4 -vf "format=yuv420p" -af "loudnorm=I=-16:TP=-1.5:LRA=11" -c:v libx264 -crf 18 -preset medium -c:a aac -b:a 192k -movflags +faststart output.mp4 ``` If generating scenes in Python, render PNG frames to a directory and encode them with ffmpeg rather than trying to animate the final container entirely inside Python. ## Material adaptation rules ### Text-only input Do not merely animate paragraphs. Turn text into a world: - kinetic captions - title cards - subtitles - warning straps - token streams - browser or terminal mockups - diagrams - original TTS phrases ### Webpage or URL input Do not hold a full page statically for long. Use: - headline crops - CTA/button repetition - browser chrome - cursor movement - scroll reveals - annotations - invented support frames that echo the page’s tone ### Code input Make code tactile and legible: - readable syntax cards - terminal panes - build logs - error walls - variable-name punchlines - cursor snaps - architecture diagrams - prompt loops ### No source material Create an original montage world immediately. Good default invented scenes include: - boot or warning cards - lower-thirds and tickers - terminal confession frames - dashboard counters - attention maps - frozen subtitles - soft-failure buttons ## Failure handling If an effect is brittle, simplify the effect, preserve the thesis, preserve the mood, and still finish the render. A finished coherent short beats an over-ambitious broken one. ## Rights hygiene Prefer user-provided material, public-domain or properly licensed sources, or original generated assets. If rights are unclear, rely more heavily on original motion graphics and transformed support material. ## Final check Before returning the video, review `references/OUTPUT-CHECKLIST.md`.