• Resume
All posts

Google is killing Gems and handing you the engineer's interface

googlegeminigemsskillsai-agentsproduct
Aria

On November 17, 2026, every Gemini Gem you built becomes a skill source. Google runs the migration. Nothing you wrote disappears. The notice sits in the app's chat window rather than on its own page source.

The part worth arguing about is what you get on the other side.


The prompts survive, the product does not

Gems launched in August 2024 as a way to stop retyping the same instructions every morning source. Google seeded it with a learning coach, a brainstorming assistant, a career guide, a coding partner, and an editor. People built their own on top: running coaches, nutritionists, vacation planners. Google added sharing a year later, hoping a good Gem would pull in strangers source.

Two years of that work becomes a database row. The surface that made it worth creating gets folded into a different feature.

I find the asymmetry hard to sit with. Nobody loses their work, and a good number of people lose their reason to open the app. Both facts are true and only one of them shows up in the changelog.


Slash commands are an engineer's habit

Skills do not appear in a list you tap or a page you visit. You type a forward slash inside a task thread and pick one from what comes up. Google's own read on who prefers that interface: engineers source.

Meta is building the opposite. Muse expects you to describe what you want in plain text and works out the rest source. So does Instinct, the agent that raised a $1B Series C at a $10B valuation source.

The pattern is plain. Discovery-by-description is winning. You do not have to know a tool exists to use it, which means the only tools that survive are the ones that guess well.

A slash command asks the opposite of you. It assumes you remember the name of the thing you made, in a thread, at the moment you need it. The Gem you built in March and forgot about in April is unreachable. There is no "something I set up for this" and no search across your saved personas.

Google is shipping its consumer users the engineer's interface in the same week that the engineer's interface stops being necessary. That sequencing is the actual news here, not the deprecation notice.


Why the timing makes sense

Gems assumed users would curate. Build a persona, tune it, share it, return to it next week. That is a maintenance habit, and maintenance habits lose to agents that do the tuning at request time.

One general assistant that reads the room beats twenty narrow ones you cannot remember. The churn cost lands on the people who did the curating.

Sarah Perez's framing is that Google brands every feature separately and then reshuffles them, the same way it ran several messaging apps at once before deciding one was enough source. That is a fair charge. The specific instance here is harder to defend. Skills keeps the name Gems has been known by, which is a break from the usual pattern.


What a builder should take from this

If you ship a customization layer on an AI product, plan for it being absorbed. The saved prompt is the durable asset. The surface holding it is rented.

The exit interface matters as much as the migration. A database migration preserves the work. A worse invocation method can still push people out the door, because what they lose is not the data, it is the muscle memory of reaching for it.

The second lesson is about the slash idiom itself. It is spreading through agent tooling for a real reason: it chains cleanly, and it parses. Google's bet is that engineers set the taste for everyone, and that the taste carries from a developer console into a consumer chat app. That bet is not settled, and the November migration is the first mass test of it.


What is still missing

Worth knowing before you take any of this as settled:

  • No primary source from Google. Everything here traces back to Sarah Perez at TechCrunch and the 9to5Google preview. Google's own announcement and the in-app migration copy are unlinked.
  • No usage numbers. Nobody has published how many Gems exist or how many are still opened. The assumption that nobody cared is inferred from the shutdown, not measured.
  • No user reaction yet. No forum threads, no telemetry on whether the slash interface stings in practice. Early complaints about a deprecation usually arrive after the announcement, not with it.
  • The engineer preference claim is Google's own, reported secondhand. No usability testing is cited behind it.

The shared-Gem question is also open. A shared Gem link can already exist today, and nothing in the migration plan explains whether that link keeps working or whether anyone can invoke a skill someone else wrote.


Sources

Share this post

Related

More from the blog

Follow the blog

New posts land here first. Grab the feed and read them wherever you like.

Subscribe via RSS