Google swaps Gems for Skills in Gemini as agent-ready prompts spread

What happened

Google is retiring Gems in the Gemini chat app and replacing them with a feature called Skills. According to The Decoder, Skills are detailed, reusable prompts that users can invoke by typing "/", or that Gemini can run automatically when a task calls for them.

The new format is based on an open standard originally developed by Anthropic. Gems will be phased out starting in November, and existing Gems will migrate to Skills automatically.

The Decoder's headline places the change within a wider shift, linking it to OpenAI and Anthropic and to a move towards prompt formats designed for AI agents. The report itself focuses on Google's implementation and timeline.

Why it matters

The change moves a prompt from a block of text that a user has to remember and retype into a named, packaged capability that the assistant can select or execute. For organisations building internal assistants or customer-facing AI tools, this makes the prompt something closer to a managed asset. It needs an owner, a clear description, version control and testing, in the same way a process or service script would.

For users, the experience shifts from recalling instructions to recognising and trusting named capabilities. Automatic invocation also means the assistant may decide when to apply a Skill, so discoverability and transparency about what a Skill does become design concerns, not just technical ones.

The Renascence take

Most readers will see a product rename, but the more significant change is that a prompt stops being a note the user rewrites and becomes a small, shareable unit of capability that the system can choose to run. Behaviourally, people struggle with blank-page prompting and with remembering the phrasing that worked last time, so named, invocable units reduce that load. The trade-off is that the less visibly a system acts on a user's behalf, the more its description has to do the explaining. Operators should audit their prompt libraries now, document each one as they would a service blueprint, and test whether the assistant triggers it at the right moments and only then.