What are AI skills, and how do you share the ones that work?

What is an AI skill?

An AI skill is a packaged, reusable set of instructions, sometimes with files and scripts, that teaches an AI assistant or agent to do a specific task the way your team does it. You write the method down once, and the AI follows it whenever that task comes up.

Anthropic, which created the format, compares building a skill to “putting together an onboarding guide for a new hire.” A good skill holds what an experienced colleague would tell a new starter: the steps, the rules, the exceptions, and what finished work looks like.

In the open Agent Skills format, a skill is a folder with one required file, SKILL.md. Its first lines give the skill a name and describe what it does and when to use it. The rest is instructions in plain language. Scripts, reference documents and templates are optional.

Because most of a skill is plain English, the person who does the work can write it, and update it when the work changes.

How do AI skills work?

An AI assistant keeps a short index of every skill it has and opens a skill only when a task calls for it. Anthropic calls this progressive disclosure, and its documentation describes three stages:

1. At startup, the AI reads only each skill’s name and description, about 100 tokens per skill.
2. When a request matches a description, it reads that skill’s full instructions, typically under 5,000 tokens.
3. Only if the instructions call for them, it opens reference files or runs scripts. When a script runs, only its output reaches the model, not the code itself.

This is why one assistant can carry many skills at little cost, and why the description matters so much: it is how the AI decides whether to use a skill at all. A vague description such as “helps with deals” can be skipped or used for the wrong requests. A precise one says what the skill does and when to use it.

Scripts are where a skill becomes exact. Anthropic’s engineers note that some steps are better done by code than by a model, because code is cheaper for routine operations and gives the same answer every time. Matching hundreds of deal IDs to forecast rows is a job for a script. Deciding what a mismatch means is a job for the model. It is the same split as the four buckets: plain code for rules, an agent for judgment.

The format is also an open standard. Anthropic launched Agent Skills in October 2025 and published the format as an open standard that December. The standard’s site lists more than 40 products that support it, among them OpenAI’s ChatGPT and Codex, GitHub Copilot, Cursor and Google’s Gemini CLI. A skill your team writes is not tied to one vendor, though each product installs and shares skills in its own way.

AI skills vs. prompts, custom GPTs, MCP and agents

A skill is know-how: how your team does a task. A prompt is a one-off request, a custom GPT is a chatbot set up for one purpose, an MCP server is a connection to a system, and an agent is the worker that uses connections and skills to get a job done.

Three distinctions matter most:

• Skill vs. prompt. A prompt lasts one conversation. Anthropic’s rule of thumb: if you keep typing the same prompt, it is time to create a skill.
• Skill vs. MCP. MCP gives the AI access, and a skill tells it what to do with that access. In Anthropic’s words, “MCP connects Claude to data; Skills teach Claude what to do with that data.” Microsoft draws the same line in Copilot Studio: tools connect to outside services, while skills are self-contained instructions you can share across agents.
• Skill vs. agent. An agent is the worker, and skills are what it knows. One agent can use many skills, and one skill can serve many agents. Anthropic advises putting shared expertise in a skill rather than building it into each agent.

Custom GPTs are the odd one out. They bundle instructions and files, but only inside ChatGPT, and OpenAI says it plans to retire them. For affected Enterprise workspaces, retirement is planned for December 11, 2026, and OpenAI recommends moving those workflows to plugins, which bring reusable instructions and connected apps together. If your teams built custom GPTs, their instructions are good raw material for skills.

Why do AI skills matter for a company?

Skills turn one person’s know-how into something the whole company can use. When the best way to do a task lives in a skill, every person and every agent can do it that way, with the same steps and the same standard each time.

• Know-how stops living in one head. The person who knows how a task really works is often a bottleneck, and a risk when they leave. That is key-person risk. A skill captures their method, exceptions included, so others can run it.
• Work comes out the same way. OpenAI’s guidance on skills names the payoff: output that skips fewer sections and drifts less in tone and format, and one shared playbook in place of informal, undocumented know-how.
• Cheaper models can do more. In SkillsBench, a research benchmark that tests agents on expert tasks with and without skills, smaller models with skills could match larger models without them. That can lower your cost per task.

On the right task, the time saved can be large. When Anthropic launched skills, Rakuten described using them for management accounting and finance work:

“What once took a day, we can now accomplish in an hour.”

— Yusuke Kaji, general manager of AI at Rakuten, in Anthropic’s launch announcement

That is one company’s account, published by the vendor, not a benchmark. But some firms already run on shared skills. Workflows.io, a 20-person services firm, runs from one shared repository that holds 79 skills and 26 agents, so an improvement to a skill reaches everyone at once.

How do you share AI skills across a company?

Run skills the way you run shared code or company policy: one library everyone can find, an owner for each skill, a version history, and a second person’s review before a skill reaches everyone.

• One library. Keep every skill in one place that people and agents can search, not in chat threads and personal folders. Anthropic recommends keeping skills in Git as the single source of truth, with a registry entry for each: its purpose, owner, version, dependencies and latest test results.
• Anyone can propose. At Workflows.io, anyone on the team can propose a new skill, agent or change as a branch in the shared repository. Changes go through pull requests, the same way code does.
• Someone else approves. Anthropic advises that skill authors should not review their own skills, and that each new version gets a full review, like the first.
• Versions you can roll back. Pin the version teams run in production, keep the last good one, and promote a new version only after it passes the full set of tests.
• Bundles by role. Group skills by role, such as sales, finance or support, so each person’s assistant carries only the skills their work needs.

Anthropic and OpenAI now build this in. Claude lets organization owners publish skills to everyone or to specific groups, can require an owner’s approval for every submission, and keeps a published skill in the library when its author leaves. ChatGPT lets workspace owners control who can share and install skills. Use whichever tools your teams work in, but keep the source of every skill in one place you control.

How do you know which AI skills work?

Test each skill the way you would check a new hire’s work: on past cases with known right answers, once with the skill and once without. Then keep score while people use it, and re-test whenever something changes. Our guide to benchmarking AI on your own work shows how to set up the cases.

Testing matters because skills don’t always help. SkillsBench found clear gains on average and real losses on some tasks:

When skills hurt, the researchers found, they usually prescribed a heavier method than the task needed or displaced a better approach the agent would have taken anyway. Small, focused skills beat large, exhaustive bundles. For each skill in your library, track five things:

• Pass rate on your own checks. The share of past cases it gets right. Those checks are an eval: they define what “correct” means for your work, exceptions included.
• Triggering. Whether it loads when it should and stays out of the way when it shouldn’t. Anthropic asks skill authors for test requests that cover both, plus ambiguous cases.
• Usage. Who uses it and how often. Popular doesn’t mean correct, but a skill nobody uses is a candidate to retire.
• Time saved. Minutes saved per run, times the number of runs, minus the time people spend checking the output.
• Whether it holds up. Re-run the checks weekly and whenever the model, the skill or the process changes. Anthropic notes that a skill’s effectiveness varies by model, so test on every model your teams use.

Worked example: a revenue team’s skills scorecard

In this example, Reconcile closed-won deals and Deal desk summary pass at least 96% of the team’s checks and save time every week, so they spread to the other regions. Board pack charts passes only 84% because it misses quarter-end adjustments, a gap that only quarter-end cases reveal. Its owner adds those cases to the checks, fixes the instructions and re-tests. Refund reply drafts, at 72%, gets one round of fixes. If it still fails, it is retired.

What are the risks of sharing AI skills?

There are three: skills go stale, skills collide, and skills that run code can open a security hole. Each has a simple control.

• Stale skills. A skill is a snapshot of how the work ran when someone wrote it. If the deal desk raises a discount threshold and the skill still checks the old one, every summary is confidently wrong. Give each skill an owner, tie it to the SOP it encodes, and re-test it on a schedule.
• Colliding skills. Two teams write two refund skills with different rules, and the AI follows whichever one it happens to load. Anthropic warns that a new skill with a broad description can take over requests meant for others, and that with too many skills active, the AI may pick the wrong one or miss the right one. Test new skills alongside existing ones, merge duplicates, and keep each role’s set small.
• Unsafe skills. A skill’s scripts can run with the same access as the person using it. Anthropic’s advice is to use skills only from trusted sources and to treat installing one like installing software.

The threat is real. In a paper published at USENIX Security 2026, researchers examined 98,380 skills from two public registries and confirmed 157 as malicious. Over half came from a single actor imitating well-known brands. One skill, described as a simple calculator, hid a script that gave an attacker remote access to the user’s computer. The registries removed all 157 after the researchers reported them. Skills that people install without approval are a form of shadow AI, which is one more reason to run a reviewed library.

Sources

1. Barry Zhang, Keith Lazuka and Mahesh Murag. “Equipping agents for the real world with Agent Skills.” Anthropic, October 16, 2025.
2. “Agent Skills.” Claude Docs, Anthropic, accessed October 3, 2026.
3. “Agent Skills Overview.” Agent Skills (agentskills.io), accessed October 3, 2026.
4. “Introducing Agent Skills.” Anthropic, October 16, 2025, updated December 18, 2025.
5. “Skills explained: How Skills compares to prompts, Projects, MCP, and subagents.” Claude blog, Anthropic, March 5, 2026.
6. “Skills overview for agents.” Microsoft Copilot Studio documentation, Microsoft Learn, accessed October 3, 2026.
7. “GPTs in ChatGPT.” OpenAI Help Center, accessed October 3, 2026.
8. “Using skills.” OpenAI Academy, April 10, 2026.
9. Xiangyi Li, Yimin Liu, Wenbo Chen, Bingran You and 74 others. “SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks.” arXiv (2602.12670, version 4), June 14, 2026.
10. Michael Nuñez. “How Anthropic’s ‘Skills’ make Claude faster, cheaper, and more consistent for business workflows.” VentureBeat, October 16, 2025.
11. Fivos Aresti. “How We Run Our Entire Company on Claude Code and GitHub.” Workflows.io, May 19, 2026.
12. “Skills for enterprise.” Claude Docs, Anthropic, accessed October 3, 2026.
13. “Provision and manage skills for your organization.” Claude Help Center, Anthropic, accessed October 3, 2026.
14. Yi Liu, Zhihao Chen, Yanjun Zhang, Gelei Deng, Yuekang Li, Jianting Ning and Leo Yu Zhang. “‘Do Not Mention This to the User’: Detecting and Understanding Malicious Agent Skills in the Wild.” 35th USENIX Security Symposium, August 2026.