Skills teach your agent how to be a powerful video editor, at every altitude: Guides choose the route for each step; techniques get the most from each model and tool.
Official
How to add sound effects and background ambience to any video (a door creaking, birds singing, wind, a room's tone, pops and whooshes for graphics): which route to take, how to place, loop, pan and mix sounds against picture, and how to check them. Use for SFX, foley, ambience, soundscapes, background sound, room tone.
Turn a still into a moving shot, or re-animate a shot from an existing spot: image-to-video (img2vid, i2v) from a pinned first frame, first and last frames pinned, reference-to-video (ref2vid, r2v) anchored on the shot's still or composed freely from references, or a video-to-video recast that keeps a take's performance. Weighs the routes, then cuts, briefs, renders and checks the shot.
How to choose a voice for a character: a short brief, a search of the ElevenLabs shared library for voices verified on eleven_v4, 3 picked on their metadata and auditioned at once on part of the script's copy, and the person's pick, with a performance guide for every script. Only when the person wants a voice matched to a face or a detailed character does a listening model rank a longer shortlist first. A voice the person cloned themselves comes in as their own recorded audio. Use for voice casting, choosing a TTS voice, voice selection, auditions.
How to make stills and illustrations of real people and real places that a viewer recognises: list the real subjects, get photos first, write each place's facts, name everything exactly in prompts, lock location plates restyled from photos for every camera direction, stage people into them, and check every still against the photos and the facts. Use for tributes, documentaries, news explainers, sports stories, history, celebrities, public figures, famous landmarks, stadiums, arenas, city skylines and any picture of someone or somewhere real.
How to build a location (a set) before any shot stills: write its geography down, render the main image of the empty set, derive every camera angle from it, then stage the cast into cropped plates, with brand colour that holds and a review page per angle. Use for set design, environments, backgrounds, studio or game-show sets, location plates and set angles.
Make someone speak on screen in sync with a voice track: lip-sync a still to audio (a presenter or avatar, a long script in chunks), fix only the mouth of an existing take, or render a dialogue scene with several people from stills and reference audio (ref2vid with audio). Weighs cost per second, quality, length limits, cutting the line by ear, and checking sync over time.
Cut the people (and the podium, chair or desk in front of the back wall) out of a span of video as a foreground cutout with a real alpha channel, a VP9 WebM that lines up frame for frame with the source: a frame-exact segment, BiRefNet v2 video on fal for the mask, and an ffmpeg alpha merge. Use it to put titles behind people, or a presenter over graphics; never colour-key instead.
Generate realistic sound effects and loopable ambience from a text description with ElevenLabs' sound effects model through hn4a (door creaks, birdsong, rain, crowds, impacts, whooshes), then place them on picture. Falls back to the sound-effects guide's other routes when hn4a doesn't offer the model.
Speak a line or a script in one ElevenLabs voice with text-to-speech on eleven_v4, directed by bracketed audio tags like [warmly] or free-form ones like [lower, knowing]. Lossless pcm_48000 output, then -16 LUFS on the finished track. Use for voice auditions, narration, a presenter's track or re-voicing one line; for a multi-voice scene in one take use elevenlabs-text-to-dialogue.
Reframe a finished 16:9 cut to 9:16 for TikTok, Reels or Shorts with ffmpeg: a vertical window per shot that holds still, pans across the shot to follow the subject and action, or fits the whole frame with padding for end cards; then the full vertical cut over the source audio, per-shot files and a contact sheet to check framing. Restaged renders splice in at exact shot length.
Lip-sync a still to an audio clip with H3 Max image-to-video and target audio: one locked presenter or speaker still plus 5 to 15 s of speech becomes a talking head that opens and closes on the still. Use for avatar and presenter videos, a long voice track cut into chunks, or a line on a shot with no usable take; to fix only the mouth of an existing take, use heygen-v3-resync-existing-take or veed-v2-resync-existing-take instead.
Titles that react to a person's hands in real footage: a title that squashes flat under a palm pushing down, one that gets shoved off screen by a sideways push, with the hand kept in front of the type. Tracks the hand locally (free), builds a hand matte, and writes per-frame box keyframes into an hn4a-compositor spec.
Build motion graphics and title slides from a JSON layer spec and render them frame-exact to an MP4 with the audio mixed in, using hn4a-compositor (headless Chromium): headlines, kickers, lists that build, logos, hand-drawn arrows and circles, a presenter as a window, cutout or round bubble, all cued to word timings. Shows the opening scene in two or three looks right after the script, cued from word times estimated from the text, before any voice exists. Live preview page, stills, auto-proxied long-GOP video, and re-rendering only a changed span with --into. Not Remotion, no build step.
Rebuild one product screen (a SaaS table, a dashboard, an app view) from a marketing screenshot as hn4a-compositor layers, so a camera can push into it at 2x and stay crisp and every part that moves is its own layer: rows that cascade and insert, cells that type, pills that flip, counts that climb, donuts and bars that draw. Covers reading the screenshot (measuring in source pixels, sampling colours per pixel, the app's own font), the builder contract (POI plus build(group_id, cues) returning layers, and a test spec), believable data, the checks, and a brief for one subagent per screen. For a product ad, app demo, UI animation or explainer; titles and cameras over footage are hn4a-compositor-in-scene-titles.
Animate a still into a shot with Kling v3 Pro image-to-video: the first frame pinned (frame 0 holds at about 0.996 against the still), an optional pinned last frame, and characters or products held from @Element references. Smooth, photographic 1080p for close-ups where texture must hold; use for frame-accurate shot recreation. For physics-heavy motion use seedance-2-5-image-to-video, for fast cheap drafts h3-max-image-to-video.
Turn a photo of a real place into a location plate in the job's art style with nano banana 2.1 edit, keeping the photo's camera, perspective and architecture exactly, its ads and logos cleared and its real signs spelled out. Use when stills must show a real stadium, building, street or landmark from several angles: one plate per camera direction, then stage people into the plates instead of generating the place.
Cut a music video's clips onto the song's bar grid: each shot gets exactly the frames between its start and end bars so cuts never drift, optional stepping on twos (12 fps look) per shot, a start offset to skip an idle opening, a short clip holding its last frame, all joined with the song underneath, and a --check that finds every cut with ffmpeg's scene score and reports any more than a frame off. Use for every rough and final cut on a song.
Add sound effects to a motion-graphics or interactive-titles cut: read the cue times from the hn4a-compositor spec, synthesize pops, squashes, splats, thwacks and panned whooshes locally with numpy/scipy (free), and swap the spec's audio for the SFX track. Use when hn4a has no sound-effect model or the person wants SFX in place of the original audio.
Time a song's lyric sheet word by word against speech-to-text of the vocal stem (ElevenLabs STT or whisper-cli JSON): each sung line matched in order, a misheard word keeping the STT's timing under the lyric's text, a dropped word interpolated and marked ?, a clean section's timing copied onto a repeat STT garbled, and every line printed with its bar. Use to cue kinetic lyric captions and karaoke-style words on the syllable.
workflows
Whether you need a narrated slide deck, a video ad, or a fully-generated movie trailer, Hypernatural has agent workflows to get you there.
Make a fully generated animated music video from a song: map the song to its bar grid, write a treatment, pick a drawn look from style frames, cast the characters on model sheets, write the shot list on the bars, make one still and one animated clip per shot, cut every shot exactly on its bars, then add kinetic lyric captions that land on the sung syllables and pulse on the beat. For cartoon and drawn-style music videos without lip sync; for a multi-look video with live-action scenes and singing on camera, use recipe-multi-style-music-video.
Turn a Figma or FigJam board (feedback, research, a retro, a dogfood wall of screenshots and stickies) into a narrated summary video: read the board, group its items by topic and rank them, write and voice a script, rebuild the board as one cluster per topic, and fly a camera across it in time with the voice. Uses the board's own images; no generated footage.
Turn a Figma or FigJam board (feedback, research, a retro, a dogfood wall of screenshots and stickies) into a narrated summary video: read the board, group its items by topic and rank them, write and voice a script, rebuild the board as one cluster per topic, and fly a camera across it in time with the voice. Uses the board's own images; no generated footage.
A skill teaches your agent how to do one part of making a video well. Guides choose the route for a step, like which video model to use or how to lip-sync a presenter. Techniques get the most from one model or tool, like how to prompt Kling or trim a clip frame for frame.
A workflow runs a whole job, from the brief to the finished cut, and calls skills for each step along the way. A skill covers one step, so your agent can also reach for it on its own, in any job.
Install Hypernatural for Agents and describe the video you want. Your agent searches the skills itself and loads the ones the job needs, so you never have to name one. You can still ask for a specific skill by name.
Any agent that can run commands on your computer and connect over MCP. Claude Code, Claude Cowork and Codex work out of the box.
No. Skills are free. You only pay for the AI generations a job uses, like video, images and voices, and your agent shows you the price of every batch before it spends anything.
Yes. When your agent figures out a trick that works, ask it to save it as a skill. You can publish it under your own handle so anyone can install it, and you can start from a copy of any skill here.
People who use Hypernatural for Agents and publish what worked for them. Every published skill is reviewed before it appears here.
Community
© 2026 Hypernatural Systems, Inc.