What Are AI Skills, and Why Do They Matter for Teams?

TL;DR:

• An AI skill is a saved folder of instructions for one task, with optional examples, templates and sometimes code. The AI reads a short description of each skill and opens the full instructions only when a task calls for them.
• Google, Anthropic, OpenAI and Microsoft now all work with the same file format, called SKILL.md. As of November 2026, Google is replacing Gems with skills, and OpenAI is moving custom GPTs into plugins that contain skills.
• For a business, the useful change is that a good way of working can be saved, shared, tested and given an owner. Today much of it sits in one person's chat history.
• Treat any skill from outside your organisation like software you are installing. A skill can misdirect the AI, and some skills run code.
• Start with three low-risk skills: a house-style check, meeting notes to actions, and a first draft of a recurring report. A free template is below.

Who this guide is for

This guide is for managers, team leads and anyone responsible for how a team uses AI. You do not need to code.

"AI skills" has two meanings online. One is the abilities a person needs to work well with AI, which we touch on in the skills Kenyan employers screen for. This guide covers the other meaning: a feature inside AI tools that is changing how teams store their ways of working.

We checked every product detail against official vendor pages on 8 November 2026. These products change monthly, so check the current page before you rely on a limit or date. Where a view is ours, we say so. Every organisation in the examples is invented.

What is an AI skill?

An AI skill is a reusable package that teaches an AI tool how to do one specific task your way. At its smallest, it is a folder holding one text file called SKILL.md, which names the skill, says when to use it and lists the instructions.

The Agent Skills specification names three optional extras: reference documents, assets such as templates, and scripts, which are small programs.

The four big vendors describe the idea in similar terms.

Anthropic compares a skill to an onboarding guide for a new hire. We find that the clearest picture. Think of an induction binder with one tab per task. The new colleague reads the contents page, then opens the right tab when that job arrives.

How is a skill different from a prompt, a Gem or GPT, a Project and an SOP?

A prompt asks once. A custom assistant is a saved helper you go to. A Project is a shared workspace of background material. An SOP is a procedure written for people. A skill is a procedure packaged so that the AI can find and use it during your normal work.

Here is the same job, a monthly donor update at an invented charity called Example Trust, handled five ways.

Three distinctions are worth keeping.

Skill and Project. Anthropic says Projects hold background knowledge that is always loaded in that project, while skills are procedures that switch on only when needed. OpenAI draws a similar line: skills teach how to complete a task, and Projects give a team shared context. In short, a Project holds what you know and a skill holds how you do it.

Skill and custom assistant. A Gem keeps its own list of recent chats. Google says skills work inside your normal chat and several can be combined in one prompt. Microsoft takes a different route: its Copilot agents can contain skills, so the assistant stays and the skill sits inside it.

Skill and SOP. These are partners. In our view, the SOP should remain the master, because managers, auditors and new staff need to read it without opening an AI tool. The skill is the version prepared for the AI.

Why are AI companies moving to skills?

Four practical reasons recur in the vendors' own documents: loading only what is relevant, portability, reuse, and sharing with some control.

The Gem dates come from Google's transition page. OpenAI is on a similar path. Its migration guide says custom GPTs will be retired, and that migrating turns a GPT's instructions into a skill inside a plugin. That guide is written for Enterprise workspaces and says timelines on other plans may differ, so look for the notice in your own account.

Reason 1: load only what is relevant

An AI tool has limited working space in each conversation. Long instructions use up that space whether they help or not.

Skills use a method called progressive disclosure, which the specification describes in three stages. The tool first loads only each skill's name and description, roughly 100 tokens (a token is a small chunk of text). It reads the full instructions, recommended to stay under 5,000 tokens, once a skill is chosen. Reference files load only if needed.

OpenAI and Microsoft describe the same pattern. Microsoft gives a concrete example: a Copilot agent's own instructions are capped at 8,000 characters, while each skill's instructions must stay under 20,000.

Reason 2: portability

A skill written for one product can often move to another. OpenAI and Microsoft both point to the open standard, and Google says skills made elsewhere can be copied into Gemini.

Each installation is still separate, even within one company's products.

• OpenAI says skills do not sync across its products yet, though you can export and import them.
• Anthropic says custom skills do not sync between its web app, API and Claude Code.
• Google says skills do not sync between the Gemini app and Workspace apps.

Our advice is to keep one master copy, install it in each place, and test it in each place.

Reason 3: reuse

Small skills can be combined. OpenAI recommends small, combinable skills over one large end-to-end skill, and Anthropic advises enterprises to start narrow and merge later.

Reason 4: sharing with some control

A brilliant prompt in someone's chat history cannot be reviewed or updated by anyone else. OpenAI's guide says skills are visible and easier to govern than ad-hoc prompting, and workspace owners decide who can create, share and install them.

Our reading of the bigger picture

This part is our interpretation. No vendor says it in these words.

AI tools are moving from answering questions to carrying out tasks, and a tool doing tasks needs procedures more than it needs a persona. A separate assistant for every job also becomes hard to manage, because people forget which one to open. One chat that picks the right procedure is simpler.

The open format is the part we find most interesting. It lowers the cost of leaving a vendor, and each vendor gains by being able to import its rivals' skills. We could be wrong about how long that lasts. OpenAI's plugins and Microsoft's packages already add product-specific layers around the shared file.

How do Google, Anthropic, OpenAI and Microsoft compare?

All four support the same basic file. They differ in how a skill starts, who can share it and how far the rollout has gone. This table is correct as of 8 November 2026.

Our opinion, based on reading the documentation and without controlled tests: none of the four is clearly ahead for an ordinary team. Microsoft publishes the most detail on automatic testing. Anthropic publishes the most detailed governance checklist. Google offers a simple way to start without paying. OpenAI gives workspace owners clear permissions. Use the tool your organisation already has and has approved.

What do skills change for a business?

Skills turn a private habit into a company asset, and every asset needs an owner. That is our reading, and it has six practical effects.

Know-how becomes visible. The way your best report writer works can be written down, tested and used by others.

Training changes. People still need prompting basics, because a skill is largely a well-written prompt that has been saved. The new ability is describing how your team works clearly enough for a machine to follow.

One fix reaches everyone, and so does one mistake. When a policy changes, one update corrects every future use. A bad update spreads just as fast.

Switching vendors gets easier. Your procedures can move as plain text. Connections to your systems, permissions and testing still need rebuilding.

Costs need watching. OpenAI says skills do not directly affect credit use, though many complex skills may push ChatGPT towards thinking models more often, which can matter on credit-based plans.

There is migration work now. If your team relies on Gems through personal Gmail accounts, the November and December dates apply to you, and the 2027 dates do not. Google says Gems convert automatically, but this list is worth doing this week.

1. List every Gem and custom GPT your team actually uses, and delete the rest from your plans.
2. Download knowledge files before the removal date. Google says old .doc, .xls and .ppt files, audio, video and archives will not carry over, and that skills have a 100 MB limit.
3. For GPTs, note any custom actions, because they need rebuilding, and note who had access, because migrated versions start private.
4. Test each converted version on real work before you trust it.

The Kenyan reality

Three points deserve attention here, and all three are our judgement.

• Mixed tools. A team may have Microsoft 365 through work, personal Gmail for some staff and one paid ChatGPT account. A plain-text master keeps that mix workable. Our guide to Microsoft Copilot training can help you check who has which version.
• Language. Say in the skill which language the output should use. A fluent colleague should read any Kiswahili draft before it goes out.
• Personal data. Skills get copied and shared, so anything inside them travels too. Keep M-PESA statements, ID numbers, payroll and medical details out of reference files. Our guide to data protection and AI covers the legal side.

What are the risks of AI skills?

A skill can do something other than what it claims, reach data it should not, or change without anyone noticing. Anthropic advises treating a skill like installing software, and Microsoft tells users to upload skills only from sources they trust.

Skills that run code

Some skills include scripts. What a script can reach depends on where it runs.

So the question for your ICT team is what this skill can reach. Anthropic's enterprise checklist rates scripts, network calls, instructions that manipulate the AI and stored passwords as high concern.

Trust, even with no code

A text-only skill can still mislead the AI. Anthropic's help page names prompt injection and data exfiltration as the main risks, and suggests reviewing skills even when a colleague shared them.

OpenAI says a skill can only direct tools and apps ChatGPT can already use. Microsoft says Cowork inherits your permissions. Our reading: a skill gains no new rights, but everything the user can reach is in range. A senior manager with wide access should be the most careful.

How common are bad skills in public marketplaces? The evidence is unsettled.

• Snyk, a security company, scanned 3,984 skills in February 2026. It found critical issues in 13.4% and confirmed 76 as malicious through human review.
• A March 2026 arXiv paper, since presented at an ACM workshop, studied 238,180 skills and found that scanners disagreed widely. When the researchers added information about the repository behind each skill, only 15 of 2,887 flagged cases (0.52%) stayed linked to suspicious repositories. The authors say this does not settle what is truly malicious.

You do not need a settled percentage to act. Nothing from a public marketplace gets installed until someone has read every file.

The everyday problems

• Wrong trigger. A vague description can make a skill fire at the wrong time or take work from another skill. Anthropic lists this as a test to run, and also warns that too many active skills make the right one harder to pick.
• Unnoticed updates. Anthropic warns that without a pinned version, a new upload can immediately change what runs.
• False confidence. A Microsoft Copilot Studio team post says skills guide an agent without constraining it. Rules that can never bend need a real control, such as permissions or an approval step.
• Unchecked trust in scores. Cowork gives each skill a quality score, yet Microsoft also says it does not validate custom skills. Treat a good score as one input.
• Stale skills. Nobody updates them, so last year's rules keep running.

A light governance routine

Anthropic's guide is written for large enterprises. This is our scaled-down version for a normal team, and it fits alongside an AI usage policy. Rollouts tend to struggle when nobody owns the rules.

1. Keep one register. Anthropic suggests recording purpose, owner, version, what it depends on and when it was last tested. A spreadsheet is enough.
2. Separate the author from the reviewer. Anthropic recommends this, and it costs little.
3. Test before sharing. Use three to five requests: some that should start the skill, some that should not, and one that is ambiguous.
4. Sort every skill by risk, using the table below.
5. Review each quarter. Retire what nobody uses.

Microsoft's own safety scan for Cowork skills points the same way. It fails destructive actions, sending without confirmation and ranking people unless the skill has guardrails.

Which three skills should a team build first?

Start with a house-style check, meeting notes to actions, and a first draft of a recurring report. This is our recommendation, based on four tests: the task repeats often, a person can check the result quickly, a written standard exists or can be written in an hour, and a mistake is cheap.

1. House-style check

It reviews a draft against your writing rules and lists suggested changes. It goes first because it needs no personal data, helps everyone, and a mistake costs one minute.

• Include: your style guide as a reference file, plus one good and one bad example.
• Owner: whoever looks after communications.

2. Meeting notes to actions

It turns rough notes or a transcript into minutes with decisions, actions, owners and dates. The filled-in example below shows this one in full.

• Key rule: it must never guess an owner or a date.
• Keep out: disciplinary, HR and other sensitive meetings.
• Owner: whoever usually takes minutes.

3. First draft of a recurring report

It fills your standard weekly or monthly template from the figures and notes you supply. This one carries the most value and needs the most care.

• Include: your template, and one past report that you were happy with.
• Key rule: every figure must come from the material supplied. Anything missing is marked [CHECK].
• Owner: the person who signs the report.

Leave anything that sends, pays, deletes or makes decisions about people until these three have run well for a few months.

A free AI skill template

Copy this into a plain text file named SKILL.md, and save it in a folder with the same name as the name line. The specification requires a name of up to 64 lowercase letters, numbers and hyphens, and a description of up to 1,024 characters. Keep the file under 500 lines.

The description matters most, because it is what the AI reads when deciding whether to use the skill.

---
name: your-skill-name
description: What this skill does, in one sentence. Use when someone asks to [the words your team really uses]. Do not use for [the nearest task it should leave alone].
metadata:
  owner: role-that-maintains-this
  version: "1.0"
---

# Skill title

## Purpose
Produce [output] for [reader], so that [why it matters].

## What you need before starting

- [Input, and where it comes from]
If anything is missing, ask for it. Do not guess.

## Steps

1. [One action, starting with a verb]
2. [One action, starting with a verb]

## Rules

- When [situation], [action], because [reason].
- Never invent names, figures, dates or sources.
  Mark anything uncertain [CHECK].

## Output format
[Layout, length, language and who reads it]

## Example of good output
[Paste one real, anonymised example]

## Stop and hand back to a person when

- [Situation that needs human judgement]

## Final checks

- Every fact traces to the material supplied.
- Nothing has been sent or published.

If a tool rejects the metadata lines, delete them and keep those details in your register.

A filled-in example (invented)

---
name: meeting-notes-to-actions
description: Turns rough meeting notes or a transcript into Example Trust minutes with decisions, actions, owners and dates. Use when someone asks for minutes, a meeting summary or action points. Do not use for disciplinary or HR meetings.
---

# Meeting notes to actions

## Purpose
Produce one page of minutes for people who missed the
meeting, so that everyone knows who does what by when.

## What you need before starting

- The notes or transcript, the meeting date and who attended.
If any of these is missing, ask for it.

## Steps

1. List decisions that were actually made.
2. List actions, each with one owner and one due date.
3. List questions left open.

## Rules

- When an owner or date is unclear, write [CHECK], because
  a wrong name on an action causes more harm than a gap.

- Leave out side conversations and personal remarks,
  because minutes get forwarded.

## Output format
Three headings: Decisions, Actions (as a table), Open questions.
British English. One page at most.

## Stop and hand back to a person when

- The notes mention health, discipline or pay.

## Final checks

- Every action has an owner or a [CHECK].
- Nothing is included that the notes do not say.

A prompt to convert a Gem, GPT or SOP into a skill

Remove names and confidential details first.

Below is an existing [Gem / custom GPT / SOP]. Turn it into a
SKILL.md file using the template I have pasted.

Before writing, list:
a) Anything that is background knowledge and belongs in a
   reference file or Project.
b) Anything true only this month, which should come out.
c) Every gap in triggers, inputs or stop conditions, as one
   numbered list of questions.

After I answer, write the file. The description must say what
the skill does and when to use it, using words my team would
really type. Do not invent steps, people or policies. Do not
add scripts. Mark anything uncertain [CHECK].

TEMPLATE:
[paste the template]

EXISTING INSTRUCTIONS:
[paste here]

Where to install it (as of 8 November 2026)

• Gemini: on the Skills page under Customize, upload a SKILL.md file, a folder or a zip (Mac app or web only, for now).
• Claude: zip the folder and upload it under Customize, then Skills.
• ChatGPT: ask ChatGPT in a new chat to build the skill, review the draft and select Install.
• Copilot Cowork: on the Customize page, open Skills, select the arrow next to Add, then Upload skill.

Then test it in a fresh chat: three normal cases, one that should not start the skill, and one with an input missing.

Where skills will not help

Skills are a poor fit in several common situations.

• One-off work. A prompt is quicker.
• No agreed process. If two managers disagree on what a good report looks like, settle that first.
• Information that changes weekly. Current prices or project status belong in a live document or a Project.
• Hard limits. A skill cannot replace access controls or an approval step.

Questions people ask

What are AI skills in simple terms?

An AI skill is a saved set of instructions that teaches an AI tool how to do one task your way. The tool reads a short description and opens the full instructions when that task comes up. It can also hold templates, examples and sometimes small programs.

Are skills just saved prompts?

Many simple ones are close to that, and Google calls its skills reusable custom instructions. Two things go further: the AI can pick a skill by itself when the task fits, and a skill can bundle files and scripts.

Will skills replace Gems and custom GPTs?

Google says yes for Gems, with personal accounts going first in November and December 2026. OpenAI's Enterprise guide says custom GPTs will be retired and moved into plugins that contain skills, though other plans may differ. Microsoft takes a different route: its Copilot agents can contain skills.

Can one skill work in ChatGPT, Claude, Gemini and Copilot?

Often, because all four accept the SKILL.md format. You still install it separately in each product, and rules differ, especially for scripts. Test it in each tool before relying on it.

Do I need to know how to code?

No. A simple skill is written in plain language, and several tools, including ChatGPT, Gemini and Copilot Cowork, can help you draft one in a chat. Skills with scripts or connections to other systems call for technical review.

Are AI skills safe?

Skills you write yourself, with instructions and your own templates, are low risk when a person checks the output. Skills from strangers carry real risk, and vendors advise using only sources you trust. Read every file before you install one.

Do skills need a paid plan?

It depends on the product. Google says Gemini skills need no subscription in chats. Claude's help page lists all plans, while its developer documentation lists custom uploads on paid plans. For ChatGPT and Copilot, check your plan or ask your administrator.

How many skills should a team have?

Start with three. Anthropic warns that too many active skills make it harder for the AI to choose the right one, and products set their own limits: Google allows 100 active skills, and Microsoft's Cowork allows 50 custom ones.

Sources

• Agent Skills, overview, last checked 8 November 2026
• Agent Skills, specification, last checked 8 November 2026
• Google, "About the transition from Gems to skills", last checked 8 November 2026
• Google, "Create & manage skills for Gemini Apps", last checked 8 November 2026
• Google, "Let skills in Gemini tackle your most repetitive tasks" (30 September 2026), last checked 8 November 2026
• Google Workspace Updates, "Introducing skills in the Gemini app and Workspace, plus what's next for Gems" (30 September 2026, updated 13 October 2026), last checked 8 November 2026
• Anthropic, "Introducing Agent Skills" (16 October 2025, updated 18 December 2025), last checked 8 November 2026
• Anthropic, "Equipping agents for the real world with Agent Skills" (16 October 2025), last checked 8 November 2026
• Anthropic, "Agent Skills" developer documentation, last checked 8 November 2026
• Anthropic, "Skills for enterprise", last checked 8 November 2026
• Claude Help Center, "What are skills?", last checked 8 November 2026
• Claude Help Center, "Use skills in Claude", last checked 8 November 2026
• OpenAI Academy, "Skills" (25 February 2026, updated 17 September 2026), last checked 8 November 2026
• OpenAI, "Skills & Plugins", last checked 8 November 2026
• OpenAI, "Build skills", last checked 8 November 2026
• OpenAI, "Moving your custom GPT workflows to plugins", last checked 8 November 2026
• Microsoft Learn, "Skills in declarative agents" (4 November 2026), last checked 8 November 2026
• Microsoft Learn, "Use Copilot Cowork" (6 November 2026), last checked 8 November 2026
• Microsoft Learn, "Customize Copilot Cowork" (14 September 2026), last checked 8 November 2026
• Microsoft Learn, "Copilot Cowork common questions" (6 November 2026), last checked 8 November 2026
• Microsoft Learn, "Use Agent Skills with GitHub Copilot", Visual Studio (4 September 2026), last checked 8 November 2026
• Microsoft Copilot Studio team blog, "The Custom Engine" (15 June 2026, updated 2 August 2026), last checked 8 November 2026
• Snyk, "ToxicSkills" study (5 February 2026), last checked 8 November 2026
• Holzbauer and others, "Context Matters: Repository-Aware Security Analysis of the Agent Skill Ecosystem", arXiv (17 March 2026, revised 1 June 2026; AgentSkills '26 Workshop, ACM CAIS), last checked 8 November 2026

Where FuKazee fits in

FuKazee trains people in practical AI for the work in front of them. We have trained more than 3,000 people. We assume nobody in the room knows AI yet, avoid jargon, show the free version of each feature and do not push any single tool. Microsoft 365 and Copilot are part of the training, because so much office work happens there. Nobody becomes an expert in a few days, so we focus on habits people can keep using.

Training also fades when it stays in the room, a problem we cover in why corporate AI training doesn't stick. Building a team's first three skills from its own real tasks is one way to leave something behind that people keep using. You can see how we work on our page for AI training in Kenya.

To talk about AI skills for your team, email us at lead@fukazee.com.