Home › Gems vs Skills
Gems and Skills both let you save custom instructions for Gemini, and the difference is easy to miss because the interface looks nearly identical. They are not the same thing. Gems are stored by Google and disappear on November 17, 2026. Skills are files you own, stored in folders you create, and they survive. Below is the difference that matters, followed by what to do with each Gem you have.
The one-line difference: a Gem is a saved instruction that lives in Google's database. A skill is a folder of files on your machine that you upload yourself.
Because of that: your Gems stop working in November, your skills do not. Google will attempt an automatic conversion, but the backup checklist exists because that conversion is not something to rely on.
| Aspect | Gems | Skills |
|---|---|---|
| Where it lives | Google's servers, managed in the Gems interface | A folder you create, containing a SKILL.md file |
| After Nov 17, 2026 | Gone. The feature is removed | Unaffected. They are your files |
| How it is created | Typed into a form in the Gemini interface | Written as a text file and uploaded, or auto-converted by Google |
| How it activates | You open it, or it is pinned. Explicit choice | It activates on its own when the request matches its description |
| Activation limits | Not documented as a hard number | Up to 100 active at once, 100 MB of uploads total |
| Sharing | Share a Gem link inside Google | Share the folder, like any file in git |
| Version history | None you control | Whatever your own file history provides |
| Works in Workspace | No | Yes, uploaded separately. The two do not sync |
| Editable after Nov 17 | No. Also not editable after October 13, 2026 | Yes. Any time, in any text editor |
This is the part that changes how you write the instructions, and it is the reason a straight copy-paste from a Gem often produces a disappointing skill.
A Gem is opened deliberately. You decide to use it, so its instructions can be written as a general set of directions. The model already knows it is in this Gem, so the text does not need to argue for its own relevance.
A skill has to earn the trigger. Before it runs, something has to decide that this skill is relevant to the request. That decision is driven almost entirely by the description field. A skill with a vague description sits there and never fires, no matter how good the body is.
| Gem-style instructions | Skill-style instructions | |
|---|---|---|
| Opening | "You are helping me draft weekly status updates." | "Drafts weekly status updates in the user's preferred format. Use when the user asks for a weekly report, a status summary, or an update for their team." |
| Assumed context | Assumed. The user already chose this Gem | Must be stated. Nobody chose anything yet |
| Tone | Direct instructions are fine | Description states what and when; the body holds the how |
Practical consequence: when you convert, the first thing to write is not the body. It is the description. If you paste Gem instructions into the body and leave the description generic, you have built something that uploads cleanly and never fires. The not triggering page covers the diagnosis; this is the prevention.
Conversion is not lossless in one direction. These Gem capabilities have no equivalent in Skills:
| Gem capability | Status in Skills | Can you rebuild it? |
|---|---|---|
| Canvas | Not supported | Partially, using a different tool |
| Deep Research | Not supported | Yes, but it is a separate workflow |
| Video generation | Not supported | Yes, with an external tool |
| Music generation | Not supported | Yes, with an external tool |
| Guided Learning | Not supported | Partially, as written instructions |
| Google Labs pre-built Gems | No conversion path | Rewrite by hand from the instructions |
The migration risks page has detail on each of these. The short version: your instruction text is portable, the features those Gems switched on are not.
Skills gained something Gems never had: files that travel with the instructions.
A skill folder can contain reference documents, templates, images, PDFs, and scripts alongside the SKILL.md. The skill can read them. That means a Gem that said "use my brand guidelines" can become a skill that actually contains the brand guidelines.
This is the strongest argument for converting rather than mourning. A skill is not a downgrade of a Gem — for most instruction-only Gems it is an upgrade, because the instructions can now reference real files.
Highest priority. Export, convert, validate, upload, enable, test. Do these first.
Export them too. You cannot recreate what you no longer remember.
Save the text, note what the feature was doing, and plan a different approach for that part.
No conversion path. Keep the instruction text and rebuild as a skill by hand.
The backup checklist walks through the mechanics, and the conversion guide handles the file structure.
Google states that it will convert Gems automatically before shutdown. Treat that as something that happens, not something you depend on. Migrated skills arrive as disabled drafts, some capabilities are dropped, and there is no documented restore path afterwards. Export your own copy.
For anything you want to keep, yes — because you own the files and they survive the shutdown. For a throwaway prompt you will use once, a Gem was simpler and that is not a criticism of it.
For instruction-only Gems, yes, once you write a description specific enough to trigger. For Gems that relied on Canvas, Deep Research, video, or music, the capability does not carry over and you need a different approach for that part.
No, not for value. Yes, for safety. The risk is not that you lose something useful, it is that you cannot remember what was in them if you ever need it.
Yes. Skills are your files in your folders, uploaded to a Google surface. Nothing about them has an end date, and the format reference documents the limits that do apply.
Not automatically. They are separate systems with separate uploads, and they do not sync. If you use both surfaces, upload to both and expect to maintain two copies.
About ten minutes per Gem including validation. The checklist has a timing table for one, five, ten, and twenty Gems.