Gemini in Chrome can now import SKILL.md agent skills: what teams should check
Chrome 157 gradually adds open agent-skill (SKILL.md) import to Gemini in Chrome. What skills are, how they run, and the trust checks for your team.
Chrome is gaining two related but different AI features that share one name, and the newer one matters to software teams. "Skills in Chrome" are saved one-click prompts for people browsing with Gemini in Chrome. Separately, as of a gradual rollout starting in Chrome 157, Gemini in Chrome supports the open agent skills specification - importable SKILL.md packages that teams already write for other AI tools.
This guide separates the two, summarizes what Google and the Chrome for Developers blog have published, and lists what a product or engineering team should check. Sources were checked on 19 November 2026; both features are rolling out gradually, so what you see in your own browser may lag the announcements.
One name, two features
Skills in Chrome (for browser users). Announced on Google's official blog in April 2026, these let a user save a useful AI prompt and re-run it with one click - by typing a forward slash or clicking the plus button in Gemini in Chrome. A saved skill runs on the page being viewed and any other tabs the user selects. Google also publishes a library of ready-made skills, and saved skills sync across a user's signed-in Chrome desktop devices. The rollout covers Mac, Windows and ChromeOS with Chrome's language set to English-US. Google states that skills ask for confirmation before certain actions, such as adding a calendar event or sending an email.
Agent skills with SKILL.md (for teams). Announced on the Chrome for Developers blog on 17 November 2026, Gemini in Chrome supports the open agent skills specification and can import skills written for other tools that use the same format. This is the part that intersects with engineering workflows.
What an agent skill actually is
A skill is a folder. At minimum it contains a SKILL.md file with metadata and instructions for performing a task; it can also carry reference documents, templates and other files:
example-skill/
├── SKILL.md # Required: metadata + instructions
├── references/ # Optional: documentation
├── assets/ # Optional: templates, resources
└── ... # Any additional files
The SKILL.md front matter carries a name and a description that tells the model when the skill applies. Chrome's example is a release-notes writer: given a page of closed issues, it groups changes, rewrites them for the reader, and returns HTML in the team's template and style guide. Chrome supports a subset of the wider specification - the developer blog lists exactly which parts, so check it before assuming a feature from another tool works in Chrome.
How skills run
Skills work with the context of the page the user is on and any tabs they share. Invocation happens two ways: the model can apply a saved skill automatically when it judges the skill relevant to the request and page, or the user can invoke one manually with the slash command or the chat menu. Once saved, a skill does not need to be requested explicitly every time.
What teams should check
- Reuse, don't rewrite. If your team already maintains
SKILL.mdfiles for other agent tools, those packages can now be imported into Chrome. Audit which of them are useful in a browser context - release notes, spec audits against team checklists, multi-tab research briefs, drafting in a consistent voice are the use cases Google names. - Treat skill sources as trusted instructions. A skill is an instruction package executed with the user's page context. Import skills only from sources you trust, and review a skill's instructions before saving it, the same way you would review code. The risks mirror the ones in our prompt injection explainer: instructions that arrive through content should not silently gain authority, and model output should be handled as untrusted input downstream.
- No website changes are required. Skills run on the user's side, in their browser, against pages they choose. There is nothing to integrate and nothing your site must add; the practical question for a site owner is only what your own team can save as reusable workflows.
- Rollout is gradual. Google notes that if you do not see the feature yet, it has not reached your browser. Both announcements describe staged rollouts, so confirm on your own Chrome version before writing documentation that depends on it.
For the other recent Chrome changes worth checking in your own product, see our Chrome 158 notes on the Web Install API, WebHID on Android and media pseudo-classes.
Sources
- Chrome for Developers, Build and import agent skills in Gemini in Chrome (17 November 2026): https://developer.chrome.com/blog/skills-in-gemini-in-chrome
- The Keyword (Google), Turn your best AI prompts into one-click tools in Chrome (14 April 2026): https://blog.google/products-and-platforms/products/chrome/skills-in-chrome/
- Chrome Help, Customize your Gemini in Chrome experience: https://support.google.com/chrome/answer/16988996
Drafted with AI assistance from these sources. Both features are in gradual rollout; verify behavior on your own Chrome version before relying on any detail.