Gemini Skills replace Gems: a saved workflow you call by API
Google is replacing Gems with Skills, reusable instructions in the Gemini app. For a saved video recipe your backend can call, Sume Formats take a POST.

If you built a Gem in the Gemini app, Google says it will migrate it to a Skill automatically when Gems go away. If you want a saved workflow that your own backend can call, that is a different object, and on Sume it is a Format: POST /v1/formats/{handle}/{slug}/runs with an instruction and input, answered by a receipt and finished media on media.sume.com.
The Gemini facts come from the Gemini Apps release notes entry dated 2026.09.30, read 2026-10-10. That entry describes Skills as "a reusable set of instructions that tell Gemini how to handle specific tasks" and says nothing about API access, so this post does not claim Skills can or cannot be called by API; it only says the note I read is silent on it.
What the release note says
According to the entry, Skills are saved shortcuts for faster prompting. You can build one from a chat, Gemini can trigger it when a matching prompt appears, and several can be stacked together for a complex task. Skills "will soon replace Gems," and Google says it will automatically migrate your Gems into skills when Gems go away, pointing users to the help center for migration steps.
That is a description of an app feature for a person typing prompts. Nothing in the note covers a spend limit, a webhook, or a machine-readable result, which is why the comparison below leaves those cells as not stated.
What a Sume Format is
A Format is a saved production recipe for one kind of video: a house style, an output contract, and a playbook, stored as a SKILL.md body with reference files. Your backend calls it by name, Sume runs it in a fresh sandbox with the generation tools, and you read back media plus, if you ask, JSON in a schema you supplied.
The recipe is set before your instruction. The agent receives the Format first and your instruction after it, so where the two disagree, the model does what you asked. You do not resend a system prompt on every call. A run is one unattended turn, so it never stops to ask a person a question.
| Question | Gemini Skills | Sume Formats |
|---|---|---|
| What is saved | Reusable instructions | A SKILL.md recipe plus reference files, versioned on each edit |
| Who starts it | A person in chat; matching prompts can trigger it | Your backend, with a Format key carrying formats:write |
| Called by HTTP | Not stated in the note | Yes: POST /v1/formats/{handle}/{slug}/runs |
| Result | Chat reply | Receipt with output, artifacts[] and primary_output_url |
| Typed output | Not stated in the note | Optional output_schema, strict subset |
| Spend ceiling | Not stated in the note | generation_spend_cap_usd per run |
Starting from a ready-made recipe
You do not have to author a Format first. Sume ships a catalog that answers at the reserved sume handle, for example sume-product-commercial, sume-virtual-try-on and sume-video-hook. Any key with formats:write can call one, and the run, its media and its spend belong to your key. Read GET /v1/formats/sume/{slug} first: it returns the description and an io profile such as url_to_video, which tells you whether the Format takes a URL, text, an image or a product.
To change a catalog recipe, fork it in the Format library. Your copy is addressed as {your_handle}/{slug}, and you call it the same way.
What migrating a Gem does not mean
Moving a Gem to a Skill is Google's migration inside Gemini. Sume does not import Gems or Skills, and a Format is not a general chat shortcut: it is built for media production, and a run is billed against your workspace wallet within its cap. Authoring is a separate surface from calling: you write the recipe in the Agents dashboard, in chat, or over the Contents API, while the run endpoints only execute it.
Two practical limits are worth knowing before you plan around it. A Format outside your key's workspace answers 404 format_not_found, and a team-owned Format needs a key created in that workspace. Long video runs usually take 15 to 30 minutes, so design for a webhook or polling from the start.
A one-call sketch
This is the shape of a catalog call, using a Format from the docs list. Replace the slug with one you have read, and set a cap you accept:
curl -sS -X POST "https://api.sume.com/v1/formats/sume/sume-video-hook/runs" \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: gems-to-formats-001" \
-d '{"instruction":"9:16 hook for a reusable water bottle.","input":{"product_name":"Trail Bottle"},"generation_spend_cap_usd":20}'Sources
Related posts
More in Comparisons
- Gemini TTS has 30 prebuilt voices; Sume TTS takes avatar voice ids
Google's Gemini TTS page lists 30 prebuilt voices. Sume TTS takes a voice UUID or voi_ id, or an avatar whose voice.status is ready; cloning stays app-only.
- Gladia diarization vs Sume speech-to-text: what you can switch
Gladia lists speaker diarization and word timestamps on every plan. Sume STT always returns word timings but fixes diarization server-side, with no switch.
- Google Vids point-in-time editing: timed text via Sume cues
Google Vids now shows only what is on screen at the playhead. Sume has no canvas, but caption cues with start and end place timed text by API.
- Grok Imagine's 4 keyframes and 7 references vs Sume's one image
xAI's Grok Imagine 1.5 takes up to 4 keyframes and up to 7 references. Sume's grok-imagine-video-1.5 row takes one image only. Rows to use for multi-image work.
Written by Sume