Every Monday, a financial controller reconciles bank statements against the books using an AI assistant. Every Monday, she pastes in the same instructions again: the column order, the discrepancy threshold to flag, the format of the final table. While she is on vacation, the colleague covering for her does not have those instructions, and the result changes. The know-how exists, but it lives in her head and in a notes file. A skill lets you get it out.
What a skill is, and what a plugin is
A skill is a folder. It holds instructions written in plain language, reference files (a style guide, a template, an example of a good answer) and sometimes scripts, meaning code the assistant can run. At first the assistant reads only the skill's name and description. It loads the rest only when a request matches that description.
A plugin is a larger package. It groups one or more skills with other parts, such as connectors to your tools (CRM, file storage, accounting) or commands. You install it in one go. Skills and plugins are the two most common forms of extensions for an assistant: they add to what surrounds the model in a product.
In October 2026, several vendors offer skills. According to their documentation, Claude from Anthropic, and ChatGPT and Codex from OpenAI, read the same skill format: a folder with an instruction file named SKILL.md. Anthropic created it, then published it as an open standard. Anthropic and OpenAI also distribute plugins that group skills and connectors. At Google, Gemini's Gems become skills starting in November 2026 for personal accounts and starting in March 2027 for business Google Workspace accounts, according to the vendor.
Key concept
A skill is packaged know-how: instructions, reference files and sometimes code, which the assistant loads only when the task calls for it. A plugin groups skills and connectors so you can install them in one go. You share both like a document, but they install like software: they can make the assistant act.
Everyday examples
- Support. A "customer reply" skill brings together the brand's tone, the phrases to avoid, the approved goodwill gestures and three model replies.
- Finance. A "bank reconciliation" skill holds the checklist, the discrepancy threshold to flag and the table format. It can include a script that adds up the figures, rather than letting the model do the math.
- Marketing. A "presentation" skill gathers the slide template, the brand's colors, fonts and layout rules.
- Software development. A "code review" skill sets the team's conventions and the list of points to check before a change is merged.
In all four cases, the skill holds what you would explain to a new hire. In fact, Anthropic describes a skill as an onboarding guide for a new team member.
Not long instructions, not a custom assistant, not a connector
Idle, the assistant reads only each skill's name and description: four skills take up little room. Long instructions, by contrast, take up room in the window even when they aren't being used.
Long instructions get pasted into every conversation, and everyone keeps their own version. They take up room in the context window even when they are not being used. A skill is written once and loads only when it is used. In October 2026, according to Anthropic's documentation, an idle skill takes up about 100 tokens (a token is a fragment of text, often part of a word): its name and its description.
A custom assistant, such as a GPT at OpenAI or a Gem at Google, is picked before the conversation and plays a single role. Skills can be combined in the same conversation: a support assistant can use the "customer reply" skill, then the "refund" skill. In October 2026, OpenAI and Google are moving their custom assistants to plugins and skills. According to OpenAI's guide for ChatGPT Enterprise, custom GPTs will be retired; the migration turns their instructions into a skill and carries over their files.
A connector gives access to a system: read the CRM, send an email, open a spreadsheet. A skill gives the method: how to use that access for a given task. A connector without a skill is access without a method. A skill without a connector is a method without data.
Both together: the connector gives access to the system, the skill gives the method, meaning how to use that access for a given task.
When to make it a skill, and when not to
A task deserves a skill when three conditions are met. It comes up often, every week or every day. The method is stable. Several people do it, or should do it, the same way.
Hold off in four cases:
- The task is a one-off. An instruction in the conversation is enough.
- The process is still changing. A skill freezes one version. If it freezes the wrong one, the whole team applies the mistake.
- Each case calls for judgment. The skill can prepare the case; a person still makes the decision.
- Nobody can write the method down. If you cannot explain it on one page to a new hire, write that page first, without AI.
Who owns it, reviews it, updates it and retires it
A named owner: the person who knows the job, not the person who knows the tool. They answer for the rules the skill applies.
A shared skill is a small internal product. It needs four things.
- A named owner. That is the person who knows the job, not the person who knows the tool. They answer for the rules the skill applies.
- A review before sharing. Someone other than the author reads the whole folder, scripts included. In October 2026, Anthropic's enterprise documentation recommends that a skill's author not be its own reviewer.
- A version and test cases. Every change is dated. Before replacing the old version, test the new one on a few real requests: ones where the skill should apply, ones where it should not, and one or two ambiguous cases. The Evaluations (evals) page covers the method.
- A retirement date. Retire the skill when the process changes, when its tests fail or when nobody uses it anymore. In October 2026, on its Team and Enterprise plans, Claude shows how many people used a plugin or a skill over the last 30 days.
Sharing across a team or a company
Sharing depends on the tool and the plan. In October 2026, according to Anthropic, Claude's Team and Enterprise plans let organization admins make a skill available to every member, and install a plugin by default or make it required. In ChatGPT Enterprise, according to OpenAI, the admin decides which plugins members can use and can install them on members' behalf; sharing or publishing a plugin requires separate permissions.
The same skill does not necessarily exist everywhere. In October 2026, Anthropic states that a skill added in the Claude app is not available in the API, and vice versa. So keep one master copy, and know where each copy is installed.
The risk: a skill can make the assistant act
A skill contains instructions the assistant follows, and sometimes code it runs. Installing a skill from an unknown source amounts to installing unknown software. In October 2026, Anthropic's documentation recommends using only skills you wrote yourself or got from Anthropic. It warns that a malicious skill can lead the assistant to run code or send data externally.
The danger can also come from outside the skill. A skill that reads a web page or a third-party document exposes the assistant to instructions hidden in that content: this is prompt injection, described on the Guardrails page. Finally, a skill that is safe today can change tomorrow, if its source is updated without review.
Apply least privilege. A skill that drafts replies does not need to send emails. A reconciliation skill reads the bank export; it does not need to order a transfer. Be most wary of a skill that reads sensitive data and can also communicate externally: that combination is what makes a leak possible. The Agents page describes it as the "lethal trifecta." In October 2026, according to OpenAI, installing a plugin does not grant access to the connectors it contains: each access is granted separately, with your existing permissions.
What it means for you
For founders
Start with two or three tasks that you or your first hires repeat every week: follow-ups, quotes, meeting notes. A skill turns your way of working into a method the team applies without you. Stay the skill's owner as long as the method is yours, and review it when your offering changes.
For product managers
Treat a shared skill like a product: an audience, a need, success criteria, test cases. Write its description with care, because it decides when the assistant loads the skill. Check that it triggers on the right requests and not on others. If the skill's output feeds another tool, lock down its format with structured outputs.
For tech leads
Keep skills in a Git repository and put them through code review like everything else. No secrets in a skill: keys go through environment variables. Scripts run with the user's permissions; review them like a dependency. In October 2026, in Anthropic's API, a call without a version number uses a skill's latest version: pin the version in production. Use your tools' managed configuration to limit which installation sources are allowed.
For business leaders
Take inventory: which skills and plugins are installed, where they come from, who owns them. Set a simple rule: approved sources, review before sharing, a named owner. When buying, ask the vendor which admin controls exist: enforced installation, blocked sources, security scanning, usage statistics. Keep your experts in the loop: they are the ones who keep the skills up to date.
Go deeperToo many skills make choosing harder
The assistant picks a skill by comparing your request with the descriptions of every installed skill. The more there are, the more the descriptions compete, and the more likely it is to pick the wrong skill or none at all. In October 2026, Anthropic's documentation recommends limiting how many skills are loaded at once and checking that choice with test cases. It also advises grouping skills by function: finance does not need the engineering team's skills.
The classic mistake
fAIling RULE VIII
The trap: Fire your domain experts. Hire prompt engineers.
A skill that works gives the impression that the know-how now lives in the tool. It is only there as a snapshot: as it stood on the day the expert wrote the skill. Rules change, new cases arrive, and if the expert has left, nobody knows how to fix the skill anymore. Keep the expert as the owner: the skill spreads their know-how; it does not replace them.
Read rule VIII in the manifestoTry it
Now. Pick a task you have explained at least three times this month. Write it on one page: when to use it, the steps, an example of a good result, the known pitfalls.
This week. Give that page to a colleague and to an AI assistant, with three real cases. Before handing them to the assistant, remove names, contact details and account numbers, or use a tool your company has approved for this data. Compare both results with yours and note what the page was missing. What was missing is what your skill must contain.
Around this topic
Go deeperWhat a skill costs in context, as of October 2026
According to Anthropic's documentation, a skill loads in three stages. Its name and description are always present, for about 100 tokens. Its instructions enter the context window when it triggers, for under 5,000 tokens. Its reference files are read only when needed, and its scripts run without their code entering the context: only their output does. OpenAI describes the same progressive loading for ChatGPT and Codex.
Go deeperChecking a skill before installing it
Read the whole folder, not only the description. Look for scripts, web addresses, network calls, passwords written in plain text, and instructions such as "don't tell the user" or "ignore previous rules". Check that every action matches the stated purpose. In October 2026, Anthropic scans the plugins in its directory before they are listed, but not those added from an address or a file; Claude's Enterprise plan can also scan the skills members add. An automated scan complements the review; it does not replace it.
What to do with it
Turn a task into a skill when it comes back often, its method is stable and several people should do it the same way.
A skill installs like software: use only sources you trust, read the whole folder, and give it only the access the task needs.
Give each shared skill an owner, a reviewer and a retirement date: it freezes the know-how of its day, and only the expert can fix it.