Google is swapping out a central building-block format in Gemini Chat: the previous Gems are being replaced by so-called Skills. Behind them are detailed, reusable prompts — that is, pre-formulated work instructions for a language model that do not have to be retyped every time. They are invoked via a forward slash in the input field; alternatively, Gemini executes them independently when the task fits.

The change is more than a renaming. The skill format is based on an open standard introduced by Anthropic. Google is thus following a path that OpenAI has also taken. For companies, this creates for the first time the prospect of no longer maintaining prompt building blocks separately for each provider.

What changes for existing Gems

The Gems will be phased out gradually from November. According to Google, existing configurations will be migrated automatically to the new format. Manual rebuilding is therefore not initially planned.

Nevertheless, a closer look is worthwhile: anyone who has run Gems that have grown over a long period with little documentation should subsequently test the migrated Skills against the actual use cases. Automatic migrations generally retain the content, but not necessarily every behaviour in detail.

Why an open format makes the difference

Until now, prompt work was in effect lock-in work. Anyone who built templates for proposal texts, meeting-note summaries or code reviews maintained them in a single provider's interface. Switching or running several models in parallel meant duplicated work.

With a shared format, this shifts. Prompt building blocks become an artefact that can be treated like source code:

  • Versionable: Changes to a skill can be traced instead of disappearing in a chat window.
  • Reviewable: The business department and technical team can examine the same text before it is used productively.
  • Transferable: The same building block can in principle be used with several providers.
  • Integrable: Skills can prospectively be integrated into your own workflows instead of existing only in chat.

That does not mean that a skill delivers identical results with every model. Different models react differently to the same instruction. The format standardises the packaging, not the behaviour.

The forward slash as an operating pattern

Invoking via "/" seems like a detail, but it is a signal. Many users know this pattern from chat and document tools. It turns a free text field into a kind of command line with a callable repertoire.

For internal rollout, this is practical. A named, findable set of skills lowers the barrier far more than a collection of prompt templates in a shared document. At the same time, a curated inventory emerges: what is not on the list is not used — and what is on it should have been reviewed.

The fact that Gemini can additionally execute Skills automatically shifts responsibility. A building block that starts only on request is chosen deliberately. A building block that kicks in on its own takes effect even when nobody currently has it in mind. Wording, data access and output formats should therefore be defined cleanly.

What teams should sensibly prepare now

The switch is a good occasion to take prompt work out of the tinkering stage. A pragmatic order suggests itself:

  1. Record the inventory: Which Gems are really in use, which are dead weight? Not everything has to come along.
  2. Clarify responsibilities: Who maintains a building block in subject-matter terms, who approves it?
  3. Define storage: A central, versioned location for the prompt texts — regardless of which interface they are used in.
  4. Test after migration: Run through a few typical inputs for each important skill and compare the results against expectations.
  5. Check the data flow: What information ends up with the provider via a skill? This question becomes sharper with automatic execution than before.

Assessment

The fact that three major providers are converging on the same format reduces a concrete risk: prompt knowledge is no longer trapped in a product interface but becomes an independent, maintainable asset. For web and software projects, this means treating prompt building blocks like configuration in future — versioned, reviewed, deployable — instead of managing them as informal text snippets in Teams chats.

Anyone integrating AI functions into their own applications thereby gains flexibility in model choice without having to rewrite the business logic every time. What remains decisive, however, is the discipline behind it: an open format does not replace clear responsibilities or tests against real use cases.

Sources