What if the shift from Gems to skills isn’t a rename at all, but a quiet admission that Google no longer believes users should be talking to a persona?
That’s the question worth sitting with. Google is shutting down Gems, the custom assistant feature in the Gemini app, and replacing it with a format called skills. Existing Gems get migrated automatically starting November 17, 2026, and no user action is required before then. On the surface it reads as housekeeping. Underneath, it’s an architectural decision with real consequences for how people interact with agents.
Invocation is architecture
The single most important detail in this change is not the name. It’s the forward slash.
Gems were things you selected and then talked to. You entered a context and stayed there. Skills are invoked by typing a slash inside a chat thread and picking one from a list. That’s a different interaction model entirely, and it tells you something about where the orchestration logic now lives.
With a persona-style custom assistant, the system prompt and context wrap the whole session. Everything you say inherits that framing. With slash-invoked skills, the framing becomes a per-turn decision. You are handing the model a tool reference at the moment of use, then returning to the general thread. The routing burden moves from the container to the call.
There’s a technical logic to this. Per-turn invocation is cleaner to reason about. It composes better, at least in theory, because multiple skills can appear in one conversation without competing system prompts fighting for priority. It also maps more directly onto how tool-calling actually works under the hood, which makes it easier for Google to unify Gemini’s consumer surface with whatever agent infrastructure sits behind it.
But cleaner for the system is not the same as clearer for the person using it. And that gap is where this change gets interesting.
The discoverability tax
Slash commands carry a cost that engineers consistently underestimate: you have to know the thing exists before you can use it.
A named assistant you select from a list is visible. It sits there as an object. A skill behind a slash key is invisible until invoked, which means the user must already hold a mental model of what’s available and what it’s called. That’s fine for developers, who have lived with slash commands in chat tools for a decade. It’s a meaningfully higher barrier for someone who just wants help writing an email.
This is why the interface is being read as less consumer-friendly than simply chatting with an all-in-one agent, the way Meta’s Muse positions itself. The competitive framing matters less than the underlying tension it exposes. There are two coherent philosophies for agent interfaces right now:
- Implicit routing — one conversational surface that infers what capability you need and dispatches internally. Lower cognitive load, less user control, more failure modes you cannot see or correct.
- Explicit invocation — the user names the capability directly. Predictable, debuggable, but requires the user to carry the map.
Slash-based skills are a bet on explicit invocation. It’s a defensible bet. It’s also a bet that runs against where consumer AI has been drifting, which is toward systems that figure it out for you.
The migration question nobody has answered
Automatic migration sounds reassuring. Architecturally, it’s the part I’d want to see tested.
If Gems carried persistent instructions and attached context, and skills are invoked per turn, then migration is not a format conversion. It’s a semantic translation between two different models of persistence. Instructions written assuming they govern an entire session behave differently when they govern a single turn. A Gem that established a voice, a set of constraints, and a working context has to become something that reasserts all of that on each invocation, or the system has to carry state forward some other way.
That translation either works well or it silently degrades behavior in ways users will notice as “it got worse” without being able to say why. Google has not published details on how the mapping handles that, and the answer determines whether this migration is a non-event or a quiet regression.
Naming churn as a signal
This also fits an established pattern. Google rebrands and merges features rather than holding a product name steady. Skills arrived with Gemini Spark and are now absorbing Gems.
Renaming is cheap internally and expensive externally. Every rename resets the accumulated understanding users built, invalidates tutorials, and breaks the shared vocabulary that lets people help each other. For agent systems specifically, where the main obstacle to adoption is that people do not know what these things can do, discarding a name users had finally learned is a real cost paid against a speculative benefit.
The deeper story here is not corporate branding. It’s that we are still watching the industry work out whether agents should be places you go or tools you call. Google appears to have picked tools you call. Whether users follow is a separate question from whether the architecture is sound.
🕒 Published: