Generate vs. Edit — The Two Kinds of Video Skills, and Why You Probably Need Both
Everyone assumes "AI video" means generation. But half the video skills we've evaluated never generate a frame — they transcribe, translate, cut, and repurpose what already exists. The two disciplines have almost opposite requirements, and picking the wrong tool for your job wastes hours.
The video category holds 43 skills in our directory. Sorted by what they actually do, they fall into two camps with almost no overlap — and the confusion between them is the most common mistake we see in skill requests.
Camp 1: Generation — footage from nothing
Text-to-video models. create-video — 9.6 — HeyGen's Video Agent from a text prompt. seedance-2-0 and seedance-2-5 — 9.0 — ByteDance's cinematic clips. ltx2 — 7.5 — open-weights text-to-video.
Programmatic composition. The HyperFrames family — ten skills, 8.7–9.7 — describe video in code and render it deterministically. anything2explainer — 9.4 — turns a topic into a produced motion-graphics explainer, complete with narration and progress bar.
Avatar and face. avatar-video — 9.2 — presenter videos with avatar control. faceswap — 8.7.
Generation skills score well when they encode the difference between a prompt that works and one that doesn't. The model API is the easy part; the craft is timing, pacing, and knowing that "make it cinematic" produces mush while "dolly-in on the subject, 24fps, shallow depth" produces footage.
Camp 2: Editing — reshaping what exists
Transcription and understanding. calesthio-video-understand — 9.2 — local frame extraction and content analysis. baoyu-youtube-transcript — 9.3 — pulls transcripts and subtitles.
Cutting and repurposing. The video-recap family — video-recap, video-assemble, video-script — 8.8–9.1 — summarizes and re-edits existing footage. video-translate — 8.8 — translates and dubs into multiple languages. music-to-video — 9.4 — pairs an existing track with new visuals.
Cleanup. video-cleaning — 8.0 — denoise, stabilize, enhance existing files.
Editing skills score well when they encode pipeline discipline: never destroy the source, cut on content boundaries not arbitrary timestamps, keep the original audio track until the last step. Their craft is workflow, not prompting.
Why the split matters
The two camps fail differently, and the failure tells you which you need:
- "The output isn't what I imagined" — you're in generation territory, and the fix is prompt craft (or switching to programmatic composition for anything that must be exact).
- "It works but it's slow and I keep redoing steps" — you're in editing territory, and the fix is a pipeline skill, not a better model.
The most expensive mistake is using generation to solve an editing problem. Turning an existing 40-minute webinar into clips is an editing job; regenerating "a webinar about X" is both worse and costlier. The best-of guide for video skills separates them for exactly this reason.
The stack that works
Our highest-scoring video workflows combine both camps in sequence:
- Understand what exists (transcription, frame analysis)
- Plan the output (script, storyboard — programmatic if it must be precise)
- Generate only the missing pieces (b-roll, narration, avatars)
- Assemble and repurpose across platforms
Teams that follow the sequence report the same thing: generation is ~10% of the work and 90% of the perceived magic, but the editing skills are what make the output shippable.
The video collection gathers the top of both camps, with full scorecards. The complete video category is 43 skills deep — every one evaluated on the same six dimensions.