top of page

Gemini Skills Add Files and Sharing

30 minutes ago
11 min read
Illustrative photograph. Photo: Godfrey Atima / Pexels.

Google is adding file support and sharing to Gemini skills as its replacement for Gems takes on more everyday work. Gemini skills file support connects reusable instructions with reference material, while sharing lets colleagues reuse those instructions. The tension is practical: distributing an AI workflow does not automatically give everyone access to the knowledge behind it.

Google announced the additions on November 16, 2026, following a gradual rollout that began November 12. Separately, 9to5Google reported that the Gems shutdown would begin rolling out November 17. Those dates describe a transition already underway, rather than a launch that suddenly reached every account on the announcement date.

The update puts pressure on people who maintain recurring prompts, team guidance, and document access. It also brings Gemini closer to the familiar custom-assistant model represented by OpenAI’s GPTs. The central question is whether a shared instruction set remains useful when its supporting documents change, its audience expands, and its original author moves on.

Gemini Skills File Support Adds Context That Can Travel

Google is connecting repeatable instructions with reusable reference material, making shared skills more useful and more dependent on document management.

Skills are reusable instructions that guide Gemini’s behavior for recurring work. With the additions, those instructions can draw on material from a device, Google Drive, Gemini Notebook, and Google Photos. A user can therefore package guidance alongside the sources that give that guidance meaning.

Google’s skill sharing update establishes the basic boundaries. Uploaded files cannot exceed 100 MB per skill. Sharing and access management currently happen through the Gemini web app, with View or Edit permissions.

Recipients receive an email and can inspect a skill before installing it. Once installed, a skill can run when Gemini identifies a relevant request, or users can invoke it explicitly. That review step matters because a shared skill carries someone else’s assumptions about how work should happen.

Consider an illustrative marketing workflow. Instructions might specify the intended audience, acceptable tone, and structure of a campaign brief. A reference document supplies approved terminology and product descriptions. Without that document, the instructions describe a process while leaving the substance to the user and model.

Attaching reference material changes the unit being shared. The recipient receives more than a well-written prompt, because the skill now points toward an established body of knowledge. That makes the package more useful, but it also creates another dependency that someone must maintain.

The distinction becomes clearer when a source changes. A writing instruction can remain appropriate while its underlying product information becomes outdated. Keeping the instruction and reference material connected reduces the need to paste revised information into every conversation.

It does not eliminate editorial judgment. A document can contain contradictory statements, outdated sections, or guidance intended for a different audience. Connecting it to a skill makes it available for use; the connection itself does not establish that every passage deserves equal weight.

The announcement therefore changes what teams should evaluate. A polished demonstration shows whether a skill produces an appealing answer under favorable conditions. A working deployment must also show whether another person can understand its purpose, access its references, and recognize when its guidance stops applying.

That second test becomes more consequential as skills move between colleagues. The author often knows which document is authoritative and which exceptions require human review. A recipient needs that context made explicit, because familiarity cannot travel through a sharing invitation.

The Gems Transition Turns Missing Features Into Immediate Work

File references and sharing address practical continuity needs during the move from Gems, but feature availability alone does not establish a successful migration.

The timing gives this release more significance than a routine convenience update. People who built repeatable work around Gems need a destination for their existing instructions and supporting material. A successor becomes credible when it can support the work users already depend on.

According to 9to5Google’s November 16 Gems deprecation coverage, the shutdown rollout was due to begin November 17. That reporting provides transition context. It should not be read as evidence that every user lost access simultaneously.

The publication’s earlier Gems migration coverage also described different schedules for personal, business, and education accounts. The important implication is that the transition has multiple audiences. Organizations should establish which schedule applies before planning changes around a consumer-facing date.

For an individual, migration friction often means recreating a familiar writing or research setup. For a team, it includes understanding who uses an existing assistant, which documents support it, and who maintains those documents. A change in interface can expose responsibilities that were previously informal.

That pressure falls first on workflow owners. A team lead who distributed recurring instructions now needs to check whether recipients can use the replacement. A knowledge manager needs to establish whether the referenced handbook remains current. An administrator needs to understand how sharing interacts with existing organizational policies.

The phrase Gemini skills vs Gems can make the transition sound like a contest between product labels. A more useful comparison asks whether existing work survives the change. Does the replacement preserve the intended behavior, accessible references, and understandable ownership that made the earlier setup useful?

Those questions are particularly relevant when the original workflow contains exceptions. A customer communication assistant might follow different guidance for public announcements and confidential updates. Copying its general tone instructions would preserve only part of the process.

There is also a difference between migration completion and operational readiness. A skill can exist in the new interface before its references, audience, and expected outputs have received review. Counting recreated objects therefore tells less than checking representative work.

None of this establishes that the transition will fail. It identifies the standard Google’s additions need to meet for users with established routines. The release supplies important capabilities, while the migration gives users an immediate reason to test their limits.

For organizations, the practical consequence is a shift from collecting prompts to maintaining shared work instructions. That responsibility remains even when the product handles installation and distribution. Someone still needs to decide which workflows deserve preservation and which should be retired.

Live References Move the Maintenance Burden Upstream

Keeping reference material current shifts maintenance toward the source document, where a small edit can affect how a shared workflow behaves.

Google says Drive files and Gemini Notebooks connected to skills update in real time. This is a company statement about reference freshness, rather than an independent finding about answer quality. The distinction matters because current information and correctly interpreted information are different outcomes.

A document connection addresses a familiar problem with copied context. Once someone pastes a handbook excerpt into an assistant, that excerpt can drift from the original. A living reference reduces the need to locate and replace every copied version after an update.

However, Gemini skills file support also changes the consequences of editing the original. A routine revision becomes an input to future AI responses. The document’s owner is therefore influencing both what people read directly and what a shared assistant uses when generating work.

Imagine a team handbook that changes its escalation procedure. An onboarding skill referencing that handbook should reflect the revised process if the connection behaves as Google describes. That is useful when the revision is deliberate, approved, and clearly written.

The same arrangement becomes harder to assess when the handbook includes unfinished proposals. Human readers might recognize a draft section from surrounding conversation. A shared skill needs instructions and source organization that make the status of that material understandable.

This is why the update has consequences for document authors, not just prompt authors. Headings, effective dates, and clear distinctions between policy and commentary gain operational importance. They help readers and AI systems identify what belongs in an answer.

OpenAI’s custom GPT knowledge documentation offers a useful competitive reference. It distinguishes behavioral instructions from uploaded reference material. That separation highlights the broader product pattern: reusable AI configurations need both directions and information to perform meaningful work.

Google’s announcement emphasizes connections to its own document services. That matters for users whose working material already lives there. It does not establish that Gemini produces better answers than competing assistants, because the release provides no comparative evaluation supporting that conclusion.

The maintenance question also changes how teams should think about approval. Approving a skill’s instructions once does not establish that every future reference revision is suitable. A dependable process needs an understood relationship between the people writing instructions and the people editing sources.

That relationship can remain lightweight for ordinary work. A small team might simply name an owner and keep approved guidance separate from brainstorming. More sensitive work warrants clearer review, because the consequences of an inaccurate source differ across tasks.

The central mechanism is straightforward: shared instructions become more useful when their references stay current. The corresponding responsibility is equally concrete. Teams need confidence that “current” means the latest approved guidance, rather than merely the most recently edited text.

Sharing the Skill Does Not Share Every Source

A skill invitation and a document permission solve different problems, and users should expect both to matter.

Google explicitly says that Drive files included in a skill must also be shared with recipients. The distinction preserves a boundary between distributing instructions and granting access to underlying information. A useful workflow can therefore remain incomplete for someone who receives only the skill.

Consider an illustrative onboarding scenario. A manager shares instructions that reference a team handbook, but the handbook remains restricted to an established employee group. A new colleague receives the workflow without necessarily receiving the information that makes its answers useful.

The announcement does not specify every response Gemini will produce when a reference is unavailable. It would be premature to promise a particular warning, fallback, or refusal behavior. That uncertainty deserves testing because users need to recognize incomplete context before relying on an answer.

A failure here does not require an inaccurate model response. Confusion about why colleagues receive different results can already undermine trust. One user might have access to an updated source while another cannot open it, despite both installing what appears to be the same skill.

The separation also protects against a different mistake: assuming that workflow sharing should automatically widen access to confidential documents. A convenient invitation should not silently become authorization for unrelated material. The challenge is making the dependency understandable without weakening the underlying boundary.

For organizational accounts, Google says skill sharing follows existing Drive sharing settings. Sharing is enabled by default, although administrators can disable it. This places the feature within an existing policy framework rather than making every workflow owner invent a separate external-sharing policy.

That arrangement still leaves a management decision. An organization can permit external collaboration while expecting employees to choose appropriate audiences for individual documents. The existence of an allowed sharing route does not establish that every skill is suitable for every recipient.

Edit permission introduces another responsibility. Someone allowed to change a skill can influence the instructions that guide its use. Teams should therefore distinguish people who need to run a workflow from people responsible for maintaining its behavior.

The announcement also says there are no administrator controls specifically for adding files to skills. That statement should be read narrowly. It describes the controls for this feature, without establishing that every other organizational security or data policy ceases to apply.

The skeptical case is therefore more practical than dramatic. Shared skills can appear ready before their document access and ownership are understood. That gap can create inconsistent work even when each individual feature behaves as designed.

For readers comparing a personal knowledge base with shared team knowledge, the update reinforces a familiar distinction. Information useful to its original owner needs additional access and maintenance decisions before it becomes a dependable team resource.

Reusable Workflows Still Need Evidence of Reliability

Google has announced a broader workflow package, but the available evidence does not show how reliably every package performs across users.

The company’s examples include shared brand guidance, onboarding instructions, and lesson planning grounded in curriculum material. These are plausible uses because they combine repeated behavior with recurring reference needs. They are examples of intended use, rather than published evidence of measured productivity gains.

That distinction should shape how the release is assessed. A feature description establishes what users are allowed to configure. It does not establish how much time they save, how consistently the system follows instructions, or how accurately it resolves competing sources.

A marketing team, for example, might value consistent terminology more than faster drafting. Its evaluation should ask whether output follows approved language across representative requests. An appealing sample paragraph would provide less evidence than repeated results covering routine and difficult cases.

An onboarding workflow faces a different standard. Employees need current procedures and a clear indication when a question falls outside the handbook. Fluency alone contributes little if an answer hides uncertainty about the applicable policy.

These examples expose the same underlying tradeoff. Packaging instructions and references makes reuse easier, which encourages distribution. Wider distribution then increases the variety of questions, permissions, and assumptions the package encounters.

The announcement does not resolve that variety through an independent benchmark. It also does not establish universal compatibility with competing vendors’ skill systems. Similar terminology should not be treated as a guarantee that instructions, files, permissions, or behavior transfer unchanged.

For anyone searching for Gemini skills explained, that is the distinction worth retaining. A skill can standardize how an assistant approaches work. It cannot, by its existence alone, certify that the source material is complete or that every generated answer meets the organization’s requirements.

The same caution applies to automatic invocation. Allowing an assistant to recognize when instructions are relevant reduces the need for explicit selection. It also makes recognition part of the workflow’s behavior, alongside the quality of the instructions themselves.

An explicit invocation provides a useful comparison during evaluation. If a request behaves differently depending on how the skill starts, that difference deserves investigation. This is an assessment question, rather than a claim that Google’s implementation necessarily produces such differences.

Teams also need to separate source problems from instruction problems. Revising the prompt will not correct an outdated handbook. Replacing a reference will not fix instructions that ask the assistant to omit necessary qualifications.

Gemini skills vs Gems should ultimately be judged through that operational lens. The successor needs to support useful work under realistic conditions, including changes in users and sources. The new features expand what can be tested, while leaving the results of those tests open.

Three Signals Will Show Whether the Transition Holds Up

The next evidence should come from availability, document behavior, and repeat use, rather than the number of skills people create.

The first signal is whether eligible users receive the announced capabilities within the stated rollout window. Google described a gradual release beginning November 12, with up to 15 days for feature visibility. Availability during that period can therefore differ between accounts.

Broad access would strengthen the case that teams can begin evaluating the same replacement capabilities together. Persistent gaps would complicate migration planning, especially when a workflow owner receives a feature before its intended recipients. Eligibility and actual visibility should remain separate observations.

This signal also needs careful attribution. A feature missing during a gradual rollout does not automatically indicate a defect. A feature still unavailable after the expected window warrants closer examination of account eligibility, administrator settings, and rollout status.

The second signal is whether source changes and access boundaries behave predictably in representative work. Google’s freshness claim gives users a concrete expectation to examine. A revised handbook should influence subsequent relevant responses without requiring users to rebuild the surrounding workflow.

That assessment should also include recipients with different document permissions. Clear handling of unavailable material would strengthen trust in shared skills. Confident answers that obscure missing context would weaken it, even if fully authorized users receive satisfactory results.

The third signal is sustained reuse after the initial migration. A team that repeatedly relies on a maintained skill provides more meaningful evidence than a large collection of unused configurations. Reuse suggests the package solves recurring work well enough to justify ongoing ownership.

This does not require inventing adoption figures Google has not published. Organizations can observe their own recurring tasks and ask whether a shared skill remains useful after references change. Over the next few months, that practical evidence will matter more than installation counts alone.

For product owners, these signals test different parts of the promise. Availability establishes access to the capability. Predictable references establish whether the package has usable context. Continued reuse establishes whether people consider the resulting workflow worth keeping.

Together, they provide a grounded way to assess Gemini skills file support without turning a feature announcement into a performance claim. The release removes important barriers to sharing repeatable work. Its lasting value depends on how the combined system behaves after that work leaves its author’s hands.

Start with a recurring task your team already understands, then compare the replacement against the work it is supposed to preserve. Identify the instruction owner, the authoritative source, and the intended recipients before treating a shared skill as dependable. Observe what happens when the source changes and when another person uses the workflow. Keep human review proportionate to the consequences of an incorrect answer. For Gemini skills file support, the useful question is whether people can repeat good work with current, accessible information. Which existing workflow would provide the clearest evidence that the transition is delivering that result?

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page