I run a directory where every agent skill gets evaluated on six dimensions — trigger quality, structure, workflow design, content, engineering, and security. Video generation is the fastest-growing category right now. Here are the ten best, tested, with the one thing you should know about each.
Video is where agent skills went from "nice trick" to "actual production tool" this year. The current wave doesn't generate video by hallucinating pixels — it writes code that draws every frame, synthesizes voiceover, and assembles the cut. That means it's debuggable, reproducible, and free of the "same-face-different-voice" artifacts that plague pure generative video.
After evaluating the full video category — 60 skills and counting — here are the ten that actually hold up.
1. remotion-to-hyperframes — 9.7/10
Converts Remotion video code into hyperframes — an intermediate format that makes video programs portable, inspectable, and reusable.
Why it's #1: The engineering discipline is exceptional — every conversion step is verified, and the output is designed to be audited, not just played. This is the rare infrastructure skill that makes everything built on top of it better.
The one thing: You need to know what Remotion is first. If you're not already writing Remotion, start with #3 and come back to this.
2. hyperframes-cli — 9.7/10
The command-line interface for the hyperframes ecosystem — build, preview, and render video programs from your terminal.
Why: CLI tools are the glue of agent workflows, and this one is engineered like a serious dev tool: proper exit codes, verifiable renders, sensible defaults. Your agent can drive the entire video pipeline without leaving the shell.
The one thing: It's a power tool. The learning curve pays off, but for one-off videos it's overkill.
3. create-video — 9.6/10
The flagship end-to-end video creation skill — script to scenes to final cut, with code-drawn frames.
Why: This is the one to install if you're starting. It handles the full pipeline (script, visuals, timing, assembly) with verification at each stage, and its task-routing sends different video types to proven recipes.
The one thing: Defaults are opinionated. First output will look "create-video flavored" until you customize the style contract.
4. website-to-video — 9.5/10
Turns any website into a polished demo or explainer video — scrolling shots, feature callouts, smooth transitions.
Why: The most immediately practical skill in the batch. Founders and marketers need product videos constantly; this generates them from the live site with no footage required. The scene sequencing logic is genuinely smart about pacing.
The one thing: Output quality tracks the design quality of the source site. A rough site gets a rough video.
5. hyperframes-core — 9.5/10
The core engine of the hyperframes format — the specification and runtime that the CLI and converter build on.
Why: If #1 and #2 are tools, this is the workshop. For teams building a video pipeline they'll reuse across projects, core gives you the primitives with the most rigorous engineering in the category.
The one thing: Pure infrastructure. No end-user features — you build those yourself.
6. music-to-video — 9.4/10
Generates music-synced visual videos — beats drive the visuals, not the other way around.
Why: Audio-visual sync is where most AI video fails embarrassingly. This skill does beat detection, tempo mapping, and scene cutting driven by the music's actual structure. The verification loop checks that cuts land on beats.
The one thing: Style presets lean abstract/visualizer. Music videos with narrative storylines aren't its strength.
7. anything2explainer — 9.4/10
Topic in, narrated explainer video out — black-canvas motion graphics with TTS voiceover, subtitles, and a chapter progress bar. Chinese or English.
Why: The most complete single-purpose pipeline evaluated. Every frame is drawn in code with Remotion (nothing generated, nothing fake), the TTS is wired in cleanly, and the chapter bar is a professional touch. This is "Kurzgesagt-style content, from a prompt."
The one thing: Needs a TTS API key configured upfront. Five minutes of setup, but it's not zero-config.
8. hyperframes-media — 9.5/10
Media management for hyperframes pipelines — asset loading, caching, and format normalization.
Why: Boring, essential, and done well. Video pipelines live or die on asset handling; this one handles formats, caching, and validation so the creative skills don't have to.
The one thing: Only meaningful if you're using multiple hyperframes skills together.
9. motion-web — 9.3/10
Builds motion-first websites on vanilla Three.js / Canvas 2D / WebGL — real physics, automated verification.
Why: Adjacent to video but worth including: it's the best implementation of "web page as motion medium" evaluated. Physics-based animation with automated visual verification catches the "looks great in my head, broken in Safari" failure.
The one thing: It makes websites, not video files. For embedded/interactive use cases rather than upload-to-YouTube ones.
10. manim-composer — 8.4/10
Orchestrates manim (the math-animation engine behind 3Blue1Brown videos) for programmatic mathematical animations.
Why: Manim is powerful but notoriously fiddly. This skill wraps it with structured composition — scenes, timing, transitions — so your agent can produce math explainers without the manim learning curve.
The one thing: The lowest score in the top ten reflects the problem domain, not the skill. Mathematical animation is hard; expect iteration.
What the other 50 video skills taught me
Three patterns from evaluating the whole category:
Code-drawn beats model-generated. The skills that draw every frame in code (Remotion, manim, Three.js) produce reproducible, brandable, artifact-free video. The ones that call image-generation APIs for every scene produce beautiful demos that fall apart on the second render.
The pipeline skills are undervalued. Everyone installs the creative skills first, but the top of the leaderboard is infrastructure (#1, #2, #5, #8). The teams shipping real video content lean on pipeline over generation.
Verification is still the differentiator. Video skills that check their own output — cuts on beats, renders that complete, assets that load — score a full tier above those that just render and hope.
Every skill here was scored on the same six-dimension rubric, and each one's full scorecard — with written rationale and a security scan of every network call — is on the video collection page. The category is growing fast; the leaderboard updates as new skills get evaluated.
What are you using to make video with agents? The good stuff hides in small repos — I'm always looking.
Top comments (1)
Full scorecards for all 60+ video skills (six dimensions, written rationale, security scan of every network call) are on the video collection page linked above. The category is the fastest-growing on the directory right now — new entries get evaluated weekly, so the leaderboard shifts. Curious what everyone's using for TTS voiceover — that's the weakest link we keep finding.