The multimodal agent Skill Omni accepts a URL or Bilibili video, extracts key screenshots and interface states, and emits a reusable procedure file — a packaged Skill an agent can invoke.
openJiuwen's community release of Skill-Omni is not an agent that reads screenshots. It is an agent that builds Skills from screenshots, and the distinction is the whole point. The same release puts a skill-engineering workflow inside the Skills system itself, so the agent can call the tool that makes the tools.
The announcement, carried by QbitAI this week, frames Skill-Omni as "the industry's first production-engineered multimodal Skill paradigm." That wording belongs to openJiuwen and should be read as vendor language. The mechanism behind it is concrete enough to describe without buying the superlative.
Today's agent Skills — the modular, reusable procedures that platforms like Anthropic's Claude Skills standard rely on — are Markdown files. They describe a task in text. That works when the important state fits in words and breaks down for anything visual: clicking through a button on a SaaS dashboard, editing a layer in a design tool, debugging a game UI. "Click the export button" leaves the agent guessing where the export button is, what state the canvas was in, and whether the click landed.
Skill-Omni, available in the JiuwenSwarm repository under the openJiuwen-ai GitHub organization, addresses that gap with a specific shape: a skill-omni-creation skill that accepts a web URL or a Bilibili video URL and emits a multimodal Skill in return. The skill pulls key screenshots, captures interface states, and traces the operation across frames before packaging the result as an executable Skill an agent can invoke.
Skill-Omni is not a separate authoring tool a human runs to author Skills. It is a Skill the agent calls. The thing that makes skills is itself a skill. openJiuwen's Swarm Skills Hub describes this as the third step in a Skill-engineering track after earlier SwarmSkill and SwarmFlow releases, and the reason each step matters is that they all layer onto the same Skill interface instead of replacing it.
Underneath sits the Swarm Skills specification on arXiv (2605.10052), which extends the Anthropic Skills standard with explicit multi-agent roles and workflows, and adds a self-evolution loop that scores skill performance. Skill-Omni is a downstream application of that spec, not a replacement for it. The platform documentation lives at openjiuwen.com/en.
The fair comparison is to Anthropic's Markdown Skills format, which has no native channel for the before/after image that an editing task depends on. Skill-Omni's claim is that closing the loop — visual input, visual output, all routed through the same Skill interface — keeps agents from losing information at the format boundary. That is a real engineering problem, not a marketing one.
Two caveats deserve to be held. First, "production-engineered" and "industry's first" are the announcement's framing, not an independent benchmark. openJiuwen points to academic predecessors such as MMSkills and VisualSkill in the launch materials but does not publish a head-to-head, and the QbitAI writeup does not either. Second, the underlying repo and Hub are public but early-stage. The interesting open question is whether outside developers can author their own visual Skills through the same self-referential pipeline, or whether the most useful extraction patterns stay curated by openJiuwen.
The watch item for the next few weeks is the repo's pace: whether skill-omni-creation ships with reproducible traces from public sources, and whether the Hub accepts community-submitted multimodal Skills. Those two signals will tell readers whether the self-referential loop is a working mechanism or a launch graphic.