이미지: X — 모델·오픈소스
Summary
- Alibaba's video generation platform Wan has officially launched an "Agent Skills" feature
- Users can browse community-made skills or build and share their own workflows
- Examples cited include a short drama generator and a music video generator
- 기능명
- Agent Skills
- 발표 주체
- Wan 공식 계정(@Alibaba_Wan)
- 발표 시점
- 2026년 8월 14일(UTC)
- 핵심 기능
- 커뮤니티 스킬 탐색, 커스텀 워크플로우 생성·공유
- 예시 워크플로우
- Short Drama Generator, MV Generator
- UI 위치
- 좌측 메뉴 Skills 탭, 상단 Explore/Skills 서브탭
- 관련 최신 모델
- Wan3.0(30초 영상 생성, 8월 6일 퍼블릭 베타)
An era of "using" workflows, not "building" them
What if, instead of writing a prompt from scratch to make a single video, you could just grab a recipe someone else already built? Alibaba's video generation platform Wan unveiled exactly this on its official account on August 14 (UTC). It's called "Agent Skills." Users can browse community-made workflows and use them right away, or build their own combinations and share them with the world.
What's changed
Screenshots from the announcement show a new "Skills" tab added to Wan's left-side menu, alongside the existing Generate, Canvas, Playground, and Assets tabs. At the top, Explore and Skills sub-tabs sit side by side, separating the screen for browsing skills made by other users from the screen for managing skills you've registered yourself. The official announcement described it as turning "ideas into reusable creative tools, from short drama generators to music video generators."
A skill is less a single finished prompt set and more an automated procedure with fixed conditions and sequencing. For example, opening a "short drama generator" skill would presumably run through pre-arranged steps for repetitive tasks like character setup, scene transitions, and dialogue insertion, in that order. However, the specific internal logic or difficulty of building a given skill wasn't detailed in this announcement.
What Wan has been building recently
Wan previously released its video generation model Wan3.0 in public beta on August 6, touting native 30-second video generation and "reality-grade" rendering. An Omni-Reference feature that accepts text, image, audio, video, and even documents and spreadsheets as reference input was introduced at the same time. This was followed on August 11 by the open-source character animation model Wan-Animate-2, which added multi-character animation and real-time streaming generation.
The images shared with this announcement also showed the Wan3.0-related input UI. The Omni input panel supports 1080P resolution, Smart Ratio, and 30-second duration options, with feature buttons listed including Reference to Video, First&Last Frame, Video Edit, Image Generation, and Agent. The Skills feature reads as an upper layer that ties these individual generation functions together for repeatable use.
How to try it
Clicking the "Skills" tab in the left-side menu of the Wan platform brings up a list of skills registered by the community. The announcement revealed two paths — "Explore skills" to browse skills, and "Upload your own" to register your own workflow.
- Go to the Wan platform and click Skills in the left-side menu.
- In the Explore tab, browse the list of published skills, such as short drama generators and music video generators.
- Select a desired skill to move to a generation screen with that workflow already applied.
- If you've built your own combination, you can share it with other users through the Upload path.
Who can try what
The announcement didn't specify usage terms or pricing tiers by country. However, "Premium" and a separate credit indicator (✦) appear at the top of the screen, suggesting a tiered usage structure similar to Wan's existing service.
The potential uses are easy to imagine. A creator making a short web drama every week could load a short-drama skill with character setup and scene transitions already built in, and just swap out the script. A musician could feed a song and lyric concept into a music video generator skill to produce a draft. Repetitive work that once required designing a new prompt each time is reduced to a few clicks.
What this means
Competition among video generation models has recently shifted beyond resolution and length toward "who can automate repetitive work more easily." The Skills feature Wan introduced here isn't about a fight over individual model performance — it's closer to a distribution structure that lets one person's know-how in building a good prompt combination be directly reused by others. Coming after Wan3.0's native 30-second generation and Wan-Animate-2's real-time streaming animation, this latest feature suggests Wan is looking to grow not just the model's own performance, but the workflow ecosystem built on top of it.
