How do you create skills in Gemini (and reuse them in Claude, ChatGPT and Copilot)?

By Andrej Ruckij · October 29, 2026 · updated November 1, 2026

TL;DR: In Gemini, open Customize > Skills, click Create, and either let Gemini build the skill in a chat, fill in a name, description and instructions yourself, or upload a SKILL.md file (or a folder or .zip with one in it). Skills work on the free plan, on the web, the mobile app and the Mac app; you call one with / or Gemini picks it up when it fits. The same file format, Agent Skills, also runs in Claude (any plan), in ChatGPT, in Microsoft 365 Copilot agents and in GitHub Copilot, so one well-written file can travel. Two things decide whether it’s worth it. First, the skill is only as good as the know-how you put into it: in the SkillsBench study, skills written by people who know the task lifted success by about 16 points on average, while in the paper’s first version, skills the model wrote for itself did slightly worse than no skill at all. Second, a skill only helps when the assistant actually loads it: in Vercel’s tests the skill was never called in 56% of cases. For personal Google accounts, Gems turn read-only on November 17 and disappear on December 17, 2026; Workspace business accounts have until March 1, 2027.

What changed, and what’s changed since

Google is replacing Gems with skills. Skills reached the Gemini app on October 13, 2026, according to Google’s admin guide to the transition, and the dates for Gems are now fixed:

Account	Gems stop being editable	Gems stop working
Personal Google account (18+)	November 17, 2026 (read-only)	December 17, 2026
Workspace Business and Enterprise		no sooner than March 1, 2027
Workspace for Education		no sooner than June 1, 2027

Remaining Gems are converted automatically. In the Workspace version they arrive as inactive draft skills that the creator has to switch on before anyone can use them. Files over the skill size limit are dropped largest-first, and unsupported files (audio, video, old .doc/.xls/.ppt files, archives) are left behind, so Google tells personal users to download any Gem files they want to keep before December 17.

On October 22, Anthropic handed the Agent Skills format to the Agentic AI Foundation under the Linux Foundation. Engineers from Anthropic, Google, OpenAI and other companies maintain it as individuals. Anthropic counts more than 70 agent platforms using the format and says 1 million skills have been published.

That second fact is why I think this matters for a small team. Until now your “how we write proposals” lived inside one product: a Custom GPT, a Gem, a Claude Project. A skill is a folder with a text file in it. Write your process down once and you can carry it to whichever assistant your team ends up paying for.

How to create skills in Gemini

Who this is for: anyone who types the same instructions into an AI assistant more than once a week: client report formats, proposal structure, brand voice, ad-copy checks, meeting prep. Who it isn’t for: one-off tasks (a normal prompt is fine), and anyone under 18. Skills need you signed in with Keep Activity on. Google’s help page says “Skills are available without a Google AI subscription.”

About 10 minutes for the first one.

Go to gemini.google.com and click Customize in the sidebar (under More if you don’t see it), then Skills at the top.
Click Create and pick one of three routes:
- Create with Gemini: describe the task and Gemini writes the skill with you in a chat.
- Create manually: enter a name, a description and the instructions yourself.
- Upload: select a SKILL.md file, or a folder or .zip with SKILL.md in its top folder. You can also start from one of the Recommended templates, or just ask in any chat: “Create a skill based on these instructions: …”.
Add reference files with Add: from your device, Google Drive, Google Photos or a notebook. A price list, a brand guide or last month’s report all work.
Use it. Type / in a chat and pick the skill, or just ask for the task; Gemini uses switched-on skills automatically when they’re relevant. You can combine skills in one prompt, for example Google’s premade /match-my-writing-style with /prep-for-meetings.

How to check it worked: run the skill on last week’s real input and put the result next to what you actually sent. If Gemini doesn’t pick the skill up without the slash command, the description is too vague (more on that below).

The limits worth knowing, from Google’s help pages on creating skills and on the move from Gems:

Skill names are lowercase with hyphens, like client-update.
All files in one skill together must stay under 100 MB. Google’s Gems transition page adds that “only 100 can be active at a time.”
Supported files include .docx, .pdf, .md, .txt, .csv, .xlsx, .pptx, images, and code files such as .py, .sh and .js. Not supported: legacy .doc/.xls/.ppt, audio, video, archives inside the skill, and code types like .ts or .sql. Remove hidden files like .DS_Store before uploading.
“Scripts that require access to the internet are not supported.” Scripts also can’t act on external websites.
To change a skill’s files you re-upload the whole skill. Deleting a skill can’t be undone. More > Download gives you a .zip, but Drive files and notebooks aren’t included in it.
Per the same transition page, “most default tools that are available in Gems do not work with skills” for now, including Create video, Create music, Canvas, Deep research and Guided learning.
The transition page says sharing skills with others will come “In the coming weeks”.

If you’re on Google Workspace: skills in the Gemini app don’t sync with skills in Workspace apps and Workspace Studio. A skill you need in both has to be built twice. Workspace Studio flows that use the “Ask a Gem” step stop working on the same 2027 dates, so they need rebuilding too.

How to write one SKILL.md that works in every tool

About 20 minutes. This is the version I’d recommend, because you write it once.

The open specification at agentskills.io sets the rules every tool follows:

A folder named in lowercase with hyphens, containing a file called SKILL.md. The folder name and the name field must match.
name: up to 64 characters; lowercase letters, numbers and single hyphens; no hyphen at the start or end.
description: up to 1,024 characters, and it “should describe both what the skill does and when to use it.”
Optional folders: scripts/ for code, references/ for longer documents, assets/ for templates and data.
Keep SKILL.md under 500 lines and move detail into the reference files, linked one level deep.

The description is the part that decides whether your skill ever runs. Every tool loads only the name and description of each skill up front (about 100 tokens each, per the spec) and reads the rest only when the description matches the request. Anthropic’s authoring guide asks for descriptions in the third person (“Processes Excel files and generates reports”, never “I can help you…”) and with the words a user would actually type.

A worked example: the weekly client report

Here’s a skill for a job most small agencies and marketing teams do every Monday. I wrote it for this article following the spec and Anthropic’s best-practices guide; treat it as a starting template, then adjust it to your own report.

Folder: weekly-client-report/

---
name: weekly-client-report
description: Turns ad-platform exports and rough notes into the weekly client report in our house format (results table, what changed, what we do next). Use when the user uploads Meta, Google or TikTok ad exports, pastes campaign numbers or notes, or asks for a weekly report, client update or performance summary.
---

# Weekly client report

## Inputs
- One or more CSV exports from Meta Ads Manager, Google Ads or TikTok Ads Manager
- Optional: the account manager's notes for the week
- Client targets are in references/targets.md. Read it before writing.

## Steps
1. Find spend, impressions, clicks, conversions and revenue per channel. If a column is missing, write "not in export". Never estimate it.
2. Calculate CPA and ROAS per channel. Compare with last week if both weeks are in the data.
3. Flag any metric more than 20% off target in references/targets.md.
4. Write the report using the template below.
5. Check before answering: every number in the text appears in the table, and every flag has a next step.

## Template
**Week of [date]: [client name]**

| Channel | Spend | Conversions | CPA | ROAS | vs target |
|---|---|---|---|---|---|

**What changed:** 2-4 sentences, plain language, no jargon.
**What we do next:** one action per flagged metric, with an owner.
**Risks:** one line. If there are none, write "No open risks."

## Rules
- British spelling, short sentences, no exclamation marks.
- Never promise results. Write "we expect" only when the notes say so.
- If the exports cover different date ranges, stop and ask which week to report.

## Example
Input note: "Meta CPA up because new creative is still learning, pause old set Thursday"
Output line: "Meta CPA rose to €41 (target €30) while the new ads are still in the learning phase. Next: Anna pauses the old ad set on Thursday."

And references/targets.md holds one short table per client: target CPA, target ROAS, monthly budget.

Why it’s built this way, from the guide: concrete input-output examples show the model the style “more clearly than descriptions alone”; a fixed template suits tasks where the format has to be the same every time; and a check step at the end (“every number in the text appears in the table”) catches the most common error before the client does. The client targets sit in a separate file so the main skill stays short and you only edit one file when a target changes.

How to check it worked: before you write the skill, Anthropic suggests running the task without one and noting what goes wrong. Then build three test cases from real past weeks and compare the output with and without the skill. If the skill doesn’t fix those specific mistakes, it isn’t earning its place yet.

Validate the file: the spec’s reference tool checks the header for you (skills-ref validate ./weekly-client-report), if someone on the team is comfortable in a terminal. Otherwise, upload to Claude: it rejects a mismatched folder name or invalid characters with an error.

How to use the same skill in Claude, ChatGPT and Copilot

Tool	Where skills live	How to add one	Plan and price (as listed, Nov 1, 2026)	Limits that matter
Gemini app	Customize > Skills	Create with Gemini, manually, or upload SKILL.md / folder / .zip	Free works; Google AI Plus €4.99, Pro €21.99, Ultra from €99.99 a month	100 MB per skill, 100 active skills, scripts can’t use the internet
Claude	Customize > Skills	+ > + Create skill > Upload a skill (ZIP of the folder)	Free, Pro ($20/mo, or $17/mo yearly), Max (from $100/mo), Team ($25/seat monthly, $20 yearly)	Needs code execution switched on; on claude.ai skills can install packages from npm and PyPI
ChatGPT	Skills in the desktop app sidebar	Build with `@skill-creator` in ChatGPT Work, or make the folder by hand; share via plugins	I couldn’t confirm which plans include it (see below)	Invoke with `@` or let ChatGPT pick
Microsoft 365 Copilot (declarative agents)	Agent Builder > Configure > Skills > Add	Upload the complete .zip, never SKILL.md alone	Needs a qualifying Microsoft 365 Copilot licence or pay-as-you-go (Agent Builder prerequisites)	8 skills per agent, 50 MB per .zip, instructions under 20,000 characters, no network in scripts, a skill can’t be reused across agents
GitHub Copilot	`.github/skills` in a repository, or `~/.copilot/skills`	Put the folder there	Not checked for this article	Developer-facing: cloud agent, code review, CLI, VS Code agent mode

Claude, step by step

About 5 minutes.

On Free, Pro or Max: Settings > Capabilities, switch on Code execution and file creation (called Create and edit files in the newer layout). On Team or Enterprise, an owner checks Organization settings > Plugins & skills > Policy that both code execution and Skills are on.
ZIP the skill folder.
Go to Customize > Skills, click +, then + Create skill, then Upload a skill.
Switch the skill on.

How to check it worked: ask for the task without naming the skill. If Claude ignores it, Anthropic’s troubleshooting list says to check that the skill is switched on, that its description is clear, and to name the skill in your request. Upload failures usually come from a ZIP over the size limit, a folder name that doesn’t match name, a missing SKILL.md, or invalid characters in the name or description.

ChatGPT

OpenAI’s “Build skills” page says standalone skills are available in the ChatGPT desktop app, Codex CLI and the IDE extension; skills bundled into plugins also work in Chat and Work on web, desktop and mobile. You call one by typing @, or ChatGPT picks it when the task matches the description. To share a skill, OpenAI’s advice is to package it as a plugin.

I found conflicting reports on which ChatGPT plans include skills, and OpenAI’s help article and pricing page blocked my reader, so I can’t give you a confirmed plan list. Check whether Skills shows in your desktop app’s sidebar before you plan around it.

Microsoft 365 Copilot

In Copilot chat, select Agents & Skills, then New agent, and either ask Agent Builder to “add a reusable skill” in plain language or open Configure > Skills > Add and upload the .zip. Then test it on the Try it tab. Microsoft’s own example is a finance team’s quarterly business review skill that bundles instructions, chart scripts and branded slide templates.

This is where the tools differ most.

Claude Team and Enterprise: an owner can upload a skill under Organization settings > Plugins & skills and it appears for everyone. Peer-to-peer sharing is on by default on Team; shared skills are view-only for recipients, and updates reach them automatically. Publishing to the organisation’s library has three modes: Requires review, Open and Off. Team plans default to Open, which means anyone can publish without a check. Only Enterprise gets “Skill and plugin security scanning”, and Anthropic says Enterprise organisations with scanning on move from Requires review to Open by default on November 4 unless an owner has already chosen a setting. If you’re on Team, I’d set publishing to Requires review on day one.
Gemini Enterprise: skills are on by default in the Business edition. Share from the More options menu with email addresses; admins can switch off “Enable skill sharing” and “Enable skills” under team settings.
Workspace Studio: creators share a link and others get a copy; the creator’s email is visible to anyone with the link. Admins decide whether skills can be shared outside the organisation.
Gemini app (personal): no sharing yet.
Microsoft 365: skills belong to one agent and can’t be reused across agents; files keep their sensitivity labels.

The security risk with skills you didn’t write

A skill is instructions the model follows, plus scripts it may run. That makes a downloaded skill closer to installing software than to copying a prompt.

Snyk scanned 3,984 public skills from ClawHub and skills.sh in February 2026: 36.82% had at least one security flaw, 534 (13.4%) had a critical issue, and human review confirmed 76 malicious payloads built for “credential theft, backdoor installation, and data exfiltration”. Snyk’s example of the pattern is a base64-encoded command that sends the user’s AWS credentials to an attacker’s server. Snyk found prompt injection in 91% of the malicious ones.

The vendors say the same in their own docs. Anthropic names “prompt injection and data exfiltration” as the main risks and says to install skills only from trusted sources. OpenAI’s developer guide says “Any Skill should be reviewed as potentially untrusted input until validated by the developer.”

Before switching on a skill someone else wrote:

Open every file, not only SKILL.md. Look for instructions to send data anywhere, to fetch other files, or to ignore earlier instructions.
Look inside scripts/ for anything that reads credentials, calls a URL, or contains long encoded strings.
Prefer skills from the vendor itself or from a colleague you can ask.
Know where your tool runs scripts. Gemini’s and Microsoft’s sandboxes have no internet access for scripts; on claude.ai skills can install packages from npm and PyPI, so a bad script has more room there.

My read: for a small team that only uses skills it wrote itself, the risk is small. It grows the day someone installs a “100 marketing skills” pack from a directory.

When a skill is the wrong tool

I like skills, and I still think most teams will over-build them. The evidence:

A skill the model writes for itself may add little or nothing. In the current version of SkillsBench (87 tasks, 18 model setups), curated skills raised the average pass rate from 33.9% to 50.5%. In the paper’s first version, skills the models wrote for themselves scored about 1-2 points below no skill. The first version’s explanation: models “generate imprecise or incomplete procedures”. So when Gemini’s Create with Gemini writes a skill from a one-line request, my read is that the skill mostly restates what the model already knows. The value is in the rules, numbers and examples only your team has.
Long skills hurt. In the paper’s first version, comprehensive skills lowered results by 2.9 points, while detailed-but-focused and compact ones did best. In that version, skills also made 16 of 84 tasks worse, which the authors put down to conflicting guidance or extra complexity on tasks the model already handled.
Skills don’t always load. Vercel tested a documentation skill for coding agents in January 2026: “In 56% of eval cases, the skill was never invoked.” The skill matched the no-skill baseline (53%) by default and reached 79% only with explicit instructions to use it; the same docs placed in always-on context reached 100%. That was a coding test, so read it as a direction for business use, not a measurement. Vercel’s conclusion is that skills “work better for vertical, action-specific workflows that users explicitly trigger.”

So my rule of thumb for a small team:

Situation	Better choice
A one-off task	A plain prompt
Facts that should apply to every chat (who your clients are, your tone, your terms)	Always-on instructions: Claude Project instructions, ChatGPT custom instructions or project files
A repeated task with a fixed output format and team-specific rules	A skill, called by name with `/` or `@`
A Gem that relies on Deep research or Canvas	Rebuild it as a skill, but test it first: Google says most default Gem tools don’t work with skills for now
A task the assistant already does well without help	Nothing. Don’t add a skill for it

The vendor case for skills is real but self-reported. Rakuten’s AI general manager told Anthropic that skills streamline its management accounting and finance workflows, and that work that took a day now takes an hour. I found no independent study of skills on marketing or ops tasks for small teams.

What I’d do this week

Pick the one instruction you’ve retyped most often this month. Before writing anything, run it once without help and note the three things the assistant gets wrong. Write a SKILL.md that fixes exactly those three, with one real example of input and output, under 100 lines. Upload it to the two tools your team uses and call it by name on three past weeks’ inputs. If both tools get all three right, that’s your template for the next skill. If you’re on a personal Google account with Gems, download their files before December 17.

More on what a skill is and how it fits with MCP: glossary/skill, tools/claude-skills, tools/mcp. A worked example of skills in real work: How do you research competitors with Claude?

Sources

More of this every day on the channel: https://t.me/aidaily_business