Gems become Skills

The short answer

It is real, not a rumor. Google announced on 30 Sep that Gemini “skills” replace Gems. Personal accounts move in November (the in-app notice says 17 Nov); Workspace moves in March 2027, schools in June 2027.
It is the same shape RichOS already runs on — saved instructions you can call, stack, or let the model pick up on its own. The Gemini app even accepts a skill uploaded as a folder with a SKILL.md in it.
So: nothing to rebuild. Make Rich's skills fit the open SKILL.md format (one file layout that Claude, ChatGPT and Gemini all read), keep them thin, and the same skill can run anywhere.

What Google is doing

The announcement. Google's blog, 30 Sep 2026, by a Gemini app product manager. Skills let you “save your most-used instructions once and run them endlessly.”
How they work. Type / and the skill's name. Stack several at once (their example: a writing-style skill plus a brand-guidelines skill). Gemini can also draft a skill from your chats and run it by itself when a prompt matches. Skills can carry reference files (text, PDF, images). Sharing and Google Drive files are “coming weeks.”
The help page adds: you can create a skill by uploading a file or folder containing a SKILL.md; scripts that need the internet are not supported; uploads cap at 100 MB; personal account, 18+, and Keep Activity on.
Gems move automatically. No action needed; Gems keep working until they are migrated.

How it lines up

Who	What it is	How it loads
Anthropic	Agent Skills: a folder with a `SKILL.md` (name + description up top, instructions below, extra files beside it). Works in claude.ai, Claude Code and the API — but each surface keeps its own copy; they do not sync.	Name and description always (~100 tokens each); the body only when triggered; extra files only when read.
OpenAI	Skills in ChatGPT and Codex, “built on the open agent skills standard” — same `SKILL.md` folder. The older Gems-like idea, custom GPTs, are separate assistants.	Names and descriptions first; full file when picked. Called with `@` in ChatGPT, `$` in Codex, or automatically.
Google	Gemini app skills (above), and Gemini CLI agent skills, which say they are based on the same open standard.	Called with `/` or picked by Gemini. The CLI loads names first, then the body on activation, after asking you.

The shared standard is real. agentskills.io publishes the SKILL.md spec. It says Anthropic first built the format and released it as an open standard; Claude, ChatGPT & Codex, Gemini CLI, Cursor, GitHub Copilot and VS Code are among the tools it lists. The rules are small: name (lowercase-and-hyphens, at most 64 characters, same as the folder) and description (what it does and when to use it, at most 1,024). Optional: license, compatibility, metadata, allowed-tools. Keep the body under 500 lines; put the rest in references/, scripts/, assets/. AGENTS.md is a different thing: one plain Markdown file of house rules for coding agents, now run by a Linux Foundation group. It is the “global rules” layer, not a skill.

Where RichOS already matches

Global rules — CLAUDE.md: first principles and preferences, pointing at the registries that hold the facts (domains, dropped list, the principles canon). This is the AGENTS.md layer.
Initiative charter — purpose, vision, principles for each initiative. The Command task builder hands the principles to the agent as binding rules, ahead of the step.
Skill — about 62 skills, each a folder with a SKILL.md carrying just a name and description up top, backed up in ! Skills. Every folder name already matches its skill name.
Spec file — the target shape: cos-brief is a short dispatcher, and _format-spec.md's Settings block overrides it.
Task — the step, its instructions, notes and attachments, last.
Where RichOS is ahead: it says who wins a conflict (CLAUDE.md beats a skill, the spec beats the skill, the charter beats the step). Google lets you stack skills but publishes no such order.

Borrow or standardize

Check every skill against the open spec. Why: it is now the one format all three companies read. Change: run the spec's own checker (skills-ref validate) over ! Skills at each backup and fix what it flags.
Put RichOS extras under metadata:. Why: the spec allows any keys there, and odd top-level keys may be rejected elsewhere (building-an-exo carries triggers: and version: up top). Change: one convention, e.g. metadata: {spec: …, version: …, owner: richos}.
Say where a skill can run, with compatibility:. Why: 28 of the skills point at iCloud paths, so they only work on Rich's Mac; Gemini's app cannot reach them. Change: one line — “Needs Rich's Mac and iCloud” — or leave it off when the skill travels.
Keep specs in one place, ship a copy at export. Why: a dispatcher that says “read ! Systems/briefs/_format/…” breaks off the Mac. Change: the spec stays the single source where it lives; the backup step copies it into the skill's references/ (store once, derive once).
Hold the 500-line line. Why: the whole body loads every time the skill fires. Five skills run 290–600 lines (building-an-exo, util-vibe-notes, util-skill-new, text-forge, two-page-primer). Change: fold this into the planned ablation pass — move detail to references/.
Write the layer order down once. Why: stacking without an order is how instructions collide, and RichOS's order is spread across three files. Change: one short “layers and who wins” section — global rules → charter → spec → skill → task — that CLAUDE.md and the task builder both cite.

Could not confirm. A 13 Oct stop on creating or editing Gems (reported by Android Authority, via tbreak; Google's own notice does not repeat it). A cap of 100 active skills (one secondary site; Google's help page gives no number). Whether Gem tool settings and share links carry over — Google has not said. Who gets skills is also unsettled: tbreak read the help page as Pro/Ultra only, while 9to5Google reports free access.

Gems become Skills — what RichOS should borrow

The short answer

What Google is doing

Where RichOS already matches

Borrow or standardize

Sources