TL;DR. Agent skills are reusable instruction folders an AI loads when a task needs them: a SKILL.md file plus optional scripts. Anthropic opened the format in December 2025; Claude, ChatGPT, Gemini and GitHub Copilot now read it. Caveat: Snyk’s February 2026 audit found critical flaws in 13.4% of 3,984 public skills, so install only trusted ones.
Last reviewed: October 3, 2026. Reviewed quarterly.
On October 13, 2026, Google starts rolling out skills in the Gemini app, the first step in replacing Gems. That makes skills a feature that Anthropic, OpenAI, and Google all support in the same file format. A year earlier, in October 2025, “agent skills” was something only Claude users and a few developers had heard of. Monthly searches for the term went from about 70 to 5,400 in under a year (Google Ads data via DataForSEO, accessed October 2026).
Agent skills are reusable, packaged instructions that an AI assistant can find and load on demand to do one specific kind of job well. Each skill is a folder containing a file called SKILL.md, which holds a name, a description of when to use it, and the step-by-step instructions. In plain terms: a skill is the written procedure you’d hand a new hire, saved where the AI can find it.
Why agent skills matter now
Agent skills matter now because the format stopped being one company’s feature and became a shared standard in about ten months. Anthropic launched Agent Skills on October 16, 2025, published the open specification on December 18, 2025, and by October 2026 OpenAI, Google, and GitHub all document the same SKILL.md format. A skill you write once can travel between tools.
- Anthropic, October 2025 and December 2025: launched Agent Skills with pre-built document skills (PowerPoint, Excel, Word, PDF), then released the open specification at agentskills.io on December 18.
- OpenAI, 2026: its developer docs now say, “Use agent skills to extend ChatGPT and Codex with task-specific capabilities,” with skills available in the ChatGPT desktop app and Codex CLI.
- Google, September 30, 2026: its Workspace Updates post says skills are “built entirely on the open, Markdown-native SKILL.md standard.” Workspace rollout begins October 5; the Gemini app rollout begins October 13.
- Search demand: “agent skills” rose from 70 to 5,400 monthly US searches between September 2025 and August 2026, and “claude skills” peaked at 90,500 in March 2026 (Google Ads data via DataForSEO).
Here’s what that means in practice. Until now, “teach the AI how we do it here” meant pasting a long prompt into every chat, or building a one-off assistant that only worked inside one product. A skill moves that knowledge into a file you own. Switch tools and the file comes with you. We track this shift because it changes what’s worth learning: the durable skill isn’t a vendor’s menu, it’s writing a procedure an AI can follow.
How agent skills actually work
An agent skill works through progressive disclosure: the AI sees only a skill’s name and one-line description until a task matches, then reads the full instructions, then opens any extra files only if the instructions point to them. This keeps the AI’s working memory (its context window) free of instructions it doesn’t need, so you can install many skills without slowing it down.
The mechanism has three levels, per Anthropic’s engineering team and the agentskills.io specification:
- Level 1, the label. At startup the AI loads just the name and description of every installed skill, roughly 100 tokens each.
- Level 2, the instructions. When your request matches a skill’s description, the AI reads that skill’s full SKILL.md. The specification recommends keeping instructions under 5,000 tokens and the file under 500 lines.
- Level 3, the extras. If the instructions mention a template, a style guide, or a script, the AI opens that file only when it reaches that step.
The technical version: a skill is a directory with one required file. The file starts with YAML frontmatter (a few labeled lines at the top) and continues in ordinary Markdown. The specification requires two fields. The name must be lowercase letters, numbers, and hyphens, at most 64 characters, and match the folder name. The description can be up to 1,024 characters and should say both what the skill does and when to use it.
---
name: weekly-client-update
description: Drafts a one-page weekly client update from raw notes, with wins, blockers, and next steps. Use when asked for a status update, weekly report, or client recap.
---
# Weekly client update
1. Ask for the client name and the week's raw notes if they are missing.
2. Write three sections: Wins, Blockers, Next week.
3. Keep it under 200 words and plain language. No jargon.
4. Never invent numbers. If a figure is missing, write "[confirm]".
That’s a complete, working skill. Optional extras live beside it: a scripts/ folder for code the AI can run, a references/ folder for longer documents, and an assets/ folder for templates. Most skills professionals write need none of those.
Agent skills vs. the things they get confused with
Agent skills differ from custom instructions, prompts, Gems, and connectors mainly in when they load and what they hold. A skill loads on demand and holds a procedure; custom instructions are always on; a connector (like MCP) gives the AI access to a tool or data source. The table below compares them using each vendor’s own documentation.
| Agent skill | Custom instructions | Gem / custom GPT | MCP connector | |
|---|---|---|---|---|
| What it holds | A procedure for one kind of task | General preferences and rules | A saved assistant with its own instructions | Access to a tool or data source |
| When it loads | Only when a task matches | Always applied | When you open that assistant | When the AI calls the tool |
| Can stack with others | Yes, several at once | n/a (one set) | One per chat | Yes |
| Portable across tools | Yes, open SKILL.md format | No | No (Gems are being replaced by skills) | Yes, open protocol |
| Can include scripts and files | Yes | No | Files only | n/a |
| Main risk | Untrusted skills can carry hidden instructions or code | Low | Low | Over-broad access |
Two sentences from the vendors capture the split. VS Code’s documentation says custom instructions are “always applied” while skills are “task-specific, loaded on-demand.” And Anthropic has said skills can complement MCP by “teaching agents more complex workflows that involve external tools.” A connector gives the AI a hand; a skill teaches it what to do with it. If you want the connector side, see our explainer on MCP.
Where agent skills show up in real work
Agent skills show up wherever someone repeats a multi-step task the same way each time: status reports, brand-voice drafts, lesson plans, month-end checklists, intake forms. The pattern is the same across industries. A person writes the procedure once, and the AI applies it whenever a matching request comes in.
| Platform | How skills appear | Where to start |
|---|---|---|
| Claude (Anthropic) | Skills in Claude apps, Claude Code, the Agent SDK, and the API; the format’s origin | Cowork Skills & Plugins Mastery |
| ChatGPT and Codex (OpenAI) | Standalone skills in the ChatGPT desktop app and Codex CLI; skills bundled in plugins work in Chat and Work | ChatGPT Workspace Agents for Non-Engineers |
| Gemini (Google) | Settings → Skills; type / to use one; replaces Gems from November 2026 on personal accounts | Gemini Gems Are Becoming Skills |
| GitHub Copilot and VS Code | Agent Skills open standard across Copilot in VS Code, Copilot CLI, and the Copilot cloud agent | VS Code Agent Skills docs |
| Other tools | Cursor, Atlassian, Figma, and others adopted the standard, per Unite.AI (December 2025) | Check each tool’s docs |
The sharpest practical difference between vendors is access. Claude and ChatGPT skills are mostly built for people who already use agents or desktop apps. Google’s rollout puts skills in a consumer chat window for personal accounts first. Work and school accounts come later, and Google’s own pages note that skills in the Gemini app and in Workspace apps don’t sync with each other.
What this means for teachers
A skill gives a teacher a lesson-planning procedure the AI follows every time, so the structure stays consistent while the topic changes. The point is not to have the AI write lessons, but to stop re-explaining your format, your reading level, and your standards each time you open a chat.
A realistic example: a skill named plan-45-minute-lesson that asks for grade, subject, and standard, then returns a table with a 5-minute warm-up, 10-minute mini-lesson, 20-minute activity, and 10-minute exit ticket, plus one support for English learners and one extension for advanced students. One skill covers every class you teach, because the grade and subject are asked for rather than hard-coded. Google’s own help pages list “a skill can help with homework and have specific instructions not to give answers” as an example use.
The honest limit: skills write the plan, not the teaching, and they can invent standard codes. Build a “never invent standard codes” rule into the skill, and check any code against the real standards document. School-managed Google accounts get skills later than personal ones.
The next step: FindSkill’s Teachers & Educators course walks through building reusable classroom helpers with AI, with the first two lessons free. Our guide to rebuilding a Gem as a skill has the full lesson-plan example.
What this means for marketers
A skill lets a marketer lock brand voice, banned phrases, and channel formats into a file, so every draft starts on-brand instead of being fixed afterward. Stacking makes this stronger: a “LinkedIn post” skill and a “brand guidelines check” skill can both shape the same draft.
The workflow: write one skill for the format (for example, a 120-word LinkedIn post with a hook, a point, and a question) and a second for the brand rules (words you never use, claims you can’t make without a source). Invoke both on a draft. Google’s launch material describes exactly this pairing, a “weekly newsletter” skill combined with an “institutional brand guidelines” skill.
The honest limit: a skill makes output consistent, not good. If your brand rules are vague (“sound friendly”), the skill is vague too. Write concrete rules (“short sentences, no exclamation points, never say ‘synergy’”), and keep a human on anything that makes a factual claim.
The next step: Custom Instructions Mastery teaches the instruction-writing habits that make a skill reliable across ChatGPT, Claude, and Gemini. Two lessons are free.
What this means for accountants and bookkeepers
A skill gives an accountant a repeatable month-end or variance-commentary procedure, with the firm’s own checklist and wording built in. Instead of asking the AI to “review this reconciliation” fresh each time, the review follows the same steps, in the same order, with the same flags.
An example: a variance-commentary skill that takes two periods of figures, asks which variances exceed the firm’s threshold, and drafts a two-sentence explanation for each in the firm’s house style, marking any figure it can’t verify with “[confirm]”. The skill never sees more than the figures you paste or attach, and it never invents a number.
The honest limit is confidentiality. A skill is only instructions and files, but the data you run through it still goes wherever your AI tool sends data. Client financials belong in a plan and a tool your firm has approved. And because skills can include scripts, don’t install an accounting skill from a stranger.
The next step: FindSkill’s Accountants & Finance course covers AI workflows for close, reconciliation, and reporting, with two lessons free.
What this means for customer support teams
A skill gives a support team a consistent way to answer a recurring kind of ticket, such as refund requests, with the policy, the tone, and the escalation rule in one place. New agents and AI drafts then work from the same page.
An example: a refund-reply skill that checks the order date against the 30-day window written into the skill, drafts an apology and a clear next step, and flags anything over a set amount for a human. Because the policy lives in the skill and not in one person’s head, updating the policy means editing one file.
The honest limit: a skill follows the policy you wrote, so a wrong policy produces consistently wrong replies. It also doesn’t verify facts about a customer’s order. Connect it to the order system through a connector, or paste the order details, and keep a human on refunds and escalations.
The next step: the AI Customer Support course covers drafting, triage, and escalation workflows, with two lessons free.
What this means for freelancers and small business owners
A skill gives a one-person business a way to package its own methods, such as a proposal format, an onboarding checklist, or a monthly invoice reminder, so the AI reproduces them without a long setup prompt each time. It’s the closest thing to hiring an assistant who already knows how you work.
An example: a client-proposal skill with your pricing tiers, your standard terms, and your preferred structure (problem, approach, timeline, price). Ask for a proposal for a new client and the draft comes back in your format, with the placeholders left blank for the things only you know.
The honest limit: skills don’t replace judgment on price, scope, or contracts. Treat the output as a first draft, and keep contract language reviewed by a professional.
The next step: the Freelancers course shows how to build AI-assisted workflows for client work, with the first lessons free.
Common misconceptions about agent skills
Most confusion about agent skills comes from treating them as either trivial (just saved prompts) or magical (a new kind of AI). They are neither. A skill is a plain-text file that changes what an AI reads at the right moment, and its value depends on how well it’s written.
“A skill is just a saved prompt.”
Partly true, which is why it’s misleading. A saved prompt is something you paste. A skill is something the AI finds and applies by itself when it matches, can combine with other skills, and can bundle with files and scripts. Google’s help pages say Gemini “automatically recognizes when a skill is relevant.” That auto-matching, plus the standard format, is the difference.
“Skills only work in Claude.”
Not anymore. Anthropic created the format, but OpenAI’s docs describe skills in ChatGPT and Codex, Google’s Workspace post builds Gemini and Workspace skills on the same SKILL.md standard, and VS Code documents it for GitHub Copilot. Support still varies: some optional fields are experimental, and Google notes that scripts needing internet access aren’t supported in Gemini skills.
“More skills is better.”
No. Every installed skill’s label is read on every request, and a vague description makes the AI use the wrong skill or none. Google caps active skills in the Gemini app at 100 for a reason. Start with three skills you’d use weekly and write sharp descriptions that say what the skill does and when to use it.
“Skills are safe because they’re just text.”
This is the one to correct. Skills can contain instructions and scripts, so a hostile one can steer an AI or run code. Anthropic’s own guidance warns that malicious skills “may introduce vulnerabilities” or “direct Claude to exfiltrate data,” and recommends installing only from trusted sources. Snyk’s February 2026 scan of 3,984 public skills found 13.4% with at least one critical issue and 76 confirmed malicious payloads. Treat a downloaded skill like a downloaded program: read it first. Our explainer on prompt injection covers the underlying attack.
Related terms
Agent skills sit inside a cluster of terms about how AI assistants are customized, connected to tools, and kept safe. These neighbors are the closest in meaning, and each links to its own plain-language explainer so you can follow whichever thread matters most to your work.
- Claude Cowork: Anthropic’s desktop agent, where skills (including recorded ones) run multi-step work.
- Agentic AI: the broader pattern of AI that takes multi-step actions toward a goal.
- MCP: the open protocol that gives AI access to tools, which skills then teach it to use.
- Prompt injection: the attack that makes untrusted skills risky.
- Context window: the AI’s working memory, which skills protect through progressive disclosure.
- Gemini Spark: Google’s background agent, where Gemini’s skills began.
- ChatGPT Projects: OpenAI’s workspace for holding instructions and files together.
See also
This section collects the courses, glossary terms, prompt templates, and articles most closely related to agent skills, grouped by type so you can jump straight to what fits your role, your experience level, and the AI tool you already use every day.
Courses on skills and related topics
- Cowork Skills & Plugins Mastery: build custom skills and plugins for Claude Cowork
- Custom Instructions Mastery: write instructions that work across ChatGPT, Claude, and Gemini
- Google Gemini Mastery: Gemini prompting, Workspace, and reusable helpers
- Build Custom OpenClaw Skills: write secure skills from scratch using the AgentSkills spec
- Claude Cowork Essentials: delegate real file work on desktop, web, and mobile
- Prompt Engineering: the instruction-writing habits behind every good skill
- AI Fundamentals: the plain-language base for everything above
- AI Agents Deep Dive: how agents plan, use tools, and remember
- Building AI Agents & Workflows: design and deploy agents step by step
- Building Custom AI Agents (No-Code): n8n, Zapier, Make, and Lindy
- Supervising AI Agents: A Manager’s Playbook: delegate, review, and fix agent work
- ChatGPT Workspace Agents for Non-Engineers: build agents that send email and run on schedule
- Copilot Cowork: Build AI Workflows in Microsoft 365: multi-step workflows in Microsoft 365
- AI Privacy 101: the settings and habits that keep your data safe
Related terms in this glossary
- Claude Cowork: Anthropic’s agentic desktop app
- Agentic AI: AI that acts toward a goal
- MCP: the protocol for connecting AI to tools
- Prompt injection: hidden text that hijacks AI instructions
- Context window: how much an AI can hold in mind at once
- Gemini Spark: Google’s always-on agent
- ChatGPT Work: OpenAI’s agent workspace
- Copilot agent: a custom assistant built in Microsoft Copilot Studio
- Multi-agent orchestration: coordinating several agents on one task
- Computer-using agent: AI that drives a real computer
AI skills (prompt templates)
- Skill Writer: guide to structuring agent skills and their frontmatter
- Custom GPT Creator: design custom GPTs with optimized instructions
- Instructions Writer: turn any process into clear step-by-step instructions
- Workflow Automator: design automations for repetitive tasks
- Agent Guardrails & Safety: safety constraints and access controls for agents
- Agentic Task Decomposer: break complex tasks into agent-friendly subtasks
Related blog posts
- Gemini Gems Are Becoming Skills: Dates, Setup, What Breaks: the official dates and a 10-minute rebuild
- Claude Skills vs Gemini Gems vs ChatGPT GPTs: the three custom-assistant formats compared
- Custom Instructions Guide: ChatGPT, Claude & Gemini: set up custom instructions on each platform
- How to Use Claude Cowork: 12 Workflows You Can Steal: workflows you can copy today
- 7 Claude Cowork Plugins Worth Installing Day 1: which plugins to install and which to skip
- AI Skills in 2026: What “AI-Fluent” Really Means: the human skills behind AI fluency
- What Is Gemini Spark?: Google’s background agent explained
Profession deep-dives
- Learn AI for Teachers: AI workflows for educators
- Learn AI for Small Business: AI for owners and operators
- Learn AI for Freelancers: AI for independent professionals
- Learn AI for Accountants: AI for finance and accounting work
The bottom line
Agent skills are the most portable way to teach an AI how you work. The format is now shared by Anthropic, OpenAI, Google, and GitHub, which means the time you spend writing a clear procedure carries across tools. The professionals who benefit most won’t be the ones who install the most skills. They’ll be the ones who write three good ones, test them on real work, and treat any skill from a stranger as software to be read before it’s run.
Frequently asked questions
What are agent skills in simple terms? Agent skills are saved instruction folders that an AI assistant loads when a task matches. Each one is a SKILL.md file with a name, a description of when to use it, and step-by-step instructions, plus optional templates and scripts. Think of it as the written procedure you’d give a new hire, stored where the AI can find it.
Are agent skills the same as Claude Skills? Claude Skills are Anthropic’s implementation, and the one that started the format in October 2025. Agent skills now refers to the open SKILL.md standard Anthropic published on December 18, 2025, which OpenAI, Google, and GitHub Copilot also support. A skill written to the standard can move between compatible tools, though optional features vary.
What is a SKILL.md file?
A SKILL.md file is a Markdown file that begins with a few lines of YAML frontmatter, a name and a description, followed by the instructions. The name must be lowercase with hyphens and at most 64 characters. The description can run to 1,024 characters and should explain what the skill does and when to use it.
How are agent skills different from custom instructions? Custom instructions are always applied to every conversation. Agent skills load only when a task matches their description, and they can hold specific procedures, files, and scripts. VS Code’s documentation puts it as skills being “task-specific, loaded on-demand” and custom instructions being “always applied.”
Do Gemini Skills replace Gems? Yes, gradually. Google’s Workspace Updates post says skills will “eventually replace Gems,” with the Gemini app rollout beginning October 13, 2026 and Gems moving to Settings on November 17, 2026. Personal-account Gems begin migrating in November; business and education accounts keep Gems until at least March 1 and June 1, 2027.
Are third-party agent skills safe to install? Not automatically. Anthropic warns that malicious skills can introduce vulnerabilities or direct an AI to exfiltrate data, and recommends installing skills only from trusted sources. Snyk’s February 2026 audit of 3,984 public skills found critical issues in 13.4% and 76 confirmed malicious payloads. Read a skill’s files before you use it.
Do I need to code to write an agent skill? No. A basic skill is a short text file with a name, a description, and a few numbered steps. Gemini and Claude can also draft one for you from a chat description. Code only enters the picture when you add optional scripts, which most professional skills don’t need.
Sources
- Anthropic, “Equipping agents for the real world with Agent Skills,” accessed 2026-10-03. https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
- Agent Skills, “Specification,” accessed 2026-10-03. https://agentskills.io/specification
- OpenAI, “Skills (ChatGPT and Codex),” developer docs, accessed 2026-10-03. https://developers.openai.com/codex/skills
- Google Workspace Updates, “Introducing skills in the Gemini app and Workspace, plus what’s next for Gems,” September 30, 2026, accessed 2026-10-03. https://workspaceupdates.googleblog.com/2026/09/skills-gemini-app-workspace.html
- Google, “About the transition from Gems to skills,” Gemini Apps Help, accessed 2026-10-03. https://support.google.com/gemini/answer/18560919?hl=en
- Google, “Create & manage skills for Gemini Apps,” Gemini Apps Help, accessed 2026-10-03. https://support.google.com/gemini/answer/17094296?hl=en
- Visual Studio Code, “Use Agent Skills in VS Code,” accessed 2026-10-03. https://code.visualstudio.com/docs/agent-customization/agent-skills
- Snyk, “ToxicSkills: 1,467 malicious payloads in an agent skills supply chain study,” February 2026, accessed 2026-10-03. https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/
- Unite.AI, “Anthropic Opens Agent Skills Standard,” December 2025, accessed 2026-10-03. https://www.unite.ai/anthropic-opens-agent-skills-standard-continuing-its-pattern-of-building-industry-infrastructure/
- PulseMCP, “OpenAI adopts Agent Skills, Anthropic donates MCP,” December 2025, accessed 2026-10-03. https://www.pulsemcp.com/posts/openai-agent-skills-anthropic-donates-mcp-gpt-5-2-image-1-5
- Google Ads search volume and keyword difficulty via DataForSEO (US), accessed 2026-10-03. https://dataforseo.com/
