How AI Agents Stopped Forgetting: What Skills Actually Do

If you have spent real time working with an AI agent, you have probably felt this: every new session, you are re-explaining the same things. Your brand voice. Your escalation policy. Your formatting preferences. It is a bit like training a new hire from scratch every Monday morning, except the hire never remembers Friday.

Skills are the industry's answer to that problem. Here is what they actually do for an agent, and how they work under the hood.

The basic idea

A skill is a directory containing a file called SKILL.md, along with optional scripts and reference material, that an AI agent can discover and load when it is relevant to the task at hand. Instead of packing every instruction into one long prompt at the start of a conversation, you write the instructions once, save them as a skill, and the agent pulls them in automatically whenever the task calls for them.

Think of it less like a prompt and more like a reference manual sitting on a shelf. The agent does not read the whole manual every time it starts work. It glances at the spine to see what is there, and only pulls the relevant chapter down when the job in front of it actually needs it.

What skills actually do for an agent

It is worth being specific here, because "skills give an agent knowledge" undersells what is actually changing.

The anatomy of a skill

At minimum, a skill is a folder with one required file: SKILL.md. That file starts with a short block of structured metadata containing two required fields — a name and a description of what the skill does and when it should be used. Below that metadata sits the actual body of instructions, written in plain Markdown.

For anything more complex than a short set of instructions, a skill folder can also include a scripts directory holding small programs the agent can run directly, a references directory holding additional documents the agent can pull in only when a specific scenario calls for them, and an assets directory holding supporting files such as templates or examples. This keeps a skill's core instructions lean while still allowing it to hold a large amount of specialised knowledge underneath.

How an agent decides to use a skill

This is the part that makes skills efficient rather than just another way to bloat a prompt. It works in layers — a pattern known as progressive disclosure. At startup, the agent loads only the name and description of every installed skill into its own system prompt, just enough to recognise when a skill might be relevant. When a task matches, the agent loads the full body of that skill's SKILL.md into its working context. If the task goes deeper still, the agent can pull in the additional reference files or run the scripts bundled with that skill — but only the ones the current task actually needs.

The effect is that an agent can have access to a very large library of specialised knowledge without carrying all of it in every conversation. Only the relevant slice gets loaded, and only when it is needed.

Skills vs prompts vs tool connections

It helps to separate three things that often get blurred together. A prompt is what you type into a single conversation — it only applies to that conversation and has to be rewritten every time. A skill is reusable procedural knowledge, how to do something, that the agent loads automatically across any conversation where it is relevant. You write it once. A tool connection, commonly built using the Model Context Protocol, gives an agent secure access to outside systems and data: a CRM, a calendar, a codebase. It is about reaching external things, not about knowing how to use them well.

A concrete example

Imagine a skill built for handling customer support tickets. Without it, an AI agent asked to draft a reply might improvise a reasonable sounding response, but with no consistent structure, no awareness of which issues need escalating, and no fixed tone.

With a skill in place, that same agent has consistent access to the actual escalation criteria, the specific ticket categories a team uses, and the tone guidelines the brand has settled on. Nothing about the agent's underlying intelligence changed. What changed is that a piece of the team's own procedural knowledge is now something the agent can reliably draw on, rather than something a person has to restate by hand each time.

Why this matters

The practical shift is this: any process a business repeats often enough to have a right way of doing it is a candidate to become a skill rather than something re-explained in every AI conversation. Building one is less like writing a clever prompt and more like writing an onboarding document — except this one gets read in full every time it is actually needed, and never gets forgotten.