What Are AI Agent Skills?

An AI agent skill is a named, packaged capability that an AI assistant can reach for when a task calls for it — not a raw function call, and not a claim that the assistant can do anything on its own. Think of it as the difference between one employee who theoretically knows everything, and one who has a specific set of trained procedures they pull out for specific jobs: order lookups, appointment booking, answering a policy question. The assistant is the employee; skills are the procedures.

Where the term comes from

“Agent Skills” is not marketing language — it’s a specific technical format, originated by Anthropic and published as an open standard. Per Anthropic’s own documentation, a skill is a folder containing a required SKILL.md file (short metadata plus instructions) and optional bundled resources — reference documents, templates, or scripts. Anthropic’s engineering team describes building one as being “organized like an onboarding guide you’d create for a new team member”: clear, scoped instructions for one job, not a dump of everything the assistant might ever need to know.

The format has since been published as an open standard at agentskills.io and adopted well beyond Anthropic’s own products — OpenAI has confirmed its Codex and ChatGPT platforms build on “the open agent skills standard,” using the same SKILL.md structure. That cross-vendor adoption is worth knowing: this is becoming a shared convention for how AI systems package reusable capabilities, not one company’s proprietary feature, in roughly the same way “REST API” describes a pattern many companies implement rather than one company’s product.

A skill is not the same as a tool

This distinction matters, and it’s easy to blur. A tool (sometimes called a function) is a single, deterministic action with a defined input and output — check an order’s status, create a calendar event. A skill is a packaged set of instructions that tells the assistant how and when to use one or more tools to get a job done well. The tool lets the assistant act; the skill tells it what good looks like for that kind of request — when to ask a clarifying question first, what tone to use, what counts as a complete answer.

Why this distinction matters for ecommerce chatbots specifically

An ecommerce chatbot vendor that only talks about “tools” or “integrations” is describing what the assistant can technically connect to — Shopify, a payments provider, a calendar. That tells you nothing about whether the assistant handles a specific customer situation well. A vendor describing “skills” is making a claim one level up: not just that it can check an order, but that it knows how to handle the various shapes a where’s-my-order conversation actually takes — a normal delivery, a delayed one, a customer who’s already contacted support twice about it. That’s a genuinely useful distinction to ask about when evaluating any chatbot platform, not just a definitional curiosity.

How AeroChat approaches this

AeroChat doesn’t market itself as implementing Anthropic’s specific SKILL.md format, and this article isn’t claiming that it does. What’s genuinely relevant to a merchant evaluating AeroChat is the same underlying idea: AeroChat’s assistant doesn’t have one undifferentiated pile of instructions for every possible customer request. It draws on distinct, scoped capabilities for distinct jobs — checking an order’s status, booking an appointment, generating a payment link, capturing a qualified lead into a CRM — each with its own rules for what counts as a complete, correct answer and when to hand off to a person instead of guessing.

That’s the practical reason the “skill” framing is useful for a merchant, independent of the specific technical standard behind it: it’s a way of asking a vendor “what does your assistant actually know how to do well, and how does it decide which of those things applies to this conversation?” rather than accepting a vague “it’s AI, it can do anything” answer. AeroChat’s own approach to the safe-execution side of this — verifying identity, validating inputs, requiring approval for sensitive actions — is covered in full in how to make a chatbot perform actions instead of just answering questions, which this article deliberately doesn’t repeat.

How this relates to conversational vs. agentic AI

“Agent skills” describes how a capability is packaged and invoked — it is not, on its own, a claim about autonomy. A chatbot can use well-defined skills for narrow, approved actions and still be conversational AI rather than fully agentic AI, which is the distinction covered in conversational AI vs. agentic AI for ecommerce support. A skill is a unit of capability, not a permission to act without oversight — a store can adopt the skill framing for how its assistant is organised internally while still requiring a human to approve anything consequential.

How to tell if a vendor actually has this, or just says it does

Ask for specifics rather than accepting the word “skills” at face value. A genuine implementation should be able to answer: which specific customer requests does the assistant handle as a distinct, well-defined capability, versus falling back to general knowledge? What happens when a request is close to one of those capabilities but not quite a match? And can the vendor show you, concretely, what a completed action looks like — a confirmed booking, a verified order status — rather than just a plausible-sounding chat response?

Frequently asked questions

Is an agent skill the same as a plugin?

They’re related but not identical. A plugin typically adds new capabilities — new tools, new data access — to a system. A skill packages instructions for using capabilities well; it can reference existing tools rather than adding new ones.

Does every AI chatbot use agent skills?

No. Many chatbots rely purely on tool or function calling without a separate packaged-skill layer. Agent skills are a specific, named pattern that’s becoming more common, not a universal requirement for a chatbot to be useful.

Why does the packaging matter, instead of just giving the model more tools?

Loading every possible tool’s full definition upfront gets expensive and can make the assistant less reliable at picking the right one. Progressive disclosure — only loading a skill’s full detail once it’s clearly relevant — keeps the assistant focused on the actual request in front of it.

Do I need to understand SKILL.md to evaluate a chatbot vendor?

No — the technical format matters to engineers building agent platforms. As a merchant, what matters is the underlying question it represents: does this assistant have well-defined, tested behaviour for the specific situations your customers actually run into, or is it improvising every time?

The next article in this series covers how AI agent skills actually work, step by step.