What is an AI skill? How your second brain goes from knowing you to doing the work
Memory is where most people stop building, and it is where the quiet disappointment starts.
You feed an AI your notes, your calls, your calendar. It remembers everything. Then Monday comes and you still open a blank document and do the work yourself. The AI knew you. It just sat there.
That gap has a name. The memory layer is passive. The skills layer is where the AI starts doing.
I build AIOS, an AI operating system you own, structured in five layers. The first two are the constitution (the rules) and memory (the second brain). This article is about the third, and it is the one that changes how the week actually feels. Memory is what the system knows. A skill is what it does with what it knows.
Memory remembers. A skill acts.
The difference between a library and a librarian.
Memory is storage. It holds your context: seven life spheres, projects, people, decisions going back years. Useful, but inert. You still have to walk in, find the right shelf, and read it yourself.
A skill is a saved workflow the AI runs for you. A set of instructions it follows the same way every time, against your real context, to produce a real output. Not a chat you start from scratch. A procedure it already knows.
Anthropic put a name to this in 2025: Agent Skills, folders of instructions and scripts a model loads when a task calls for them, so it does not relearn the same procedure every session. Same idea, whether you build it yourself or use a tool. A repeatable thing your AI knows how to run.
A skill is a workflow your AI runs, not a prompt you retype.
The boring kind of automation, the kind that actually saves hours.
Skills are not exotic. They are the repetitive work you already do badly, because you are tired by the time you get to it.
A few that earn their place:
- The weekly review. Every Sunday the AI pulls your week, calendar, calls, what moved, what stalled, and drafts the review for you to edit.
- The proposal draft. A lead comes in. The skill reads your past proposals and the call notes, then writes a first draft in your structure and your voice.
Inbox triage belongs here too. The AI sorts what needs you today from what can wait, instead of you reading 40 emails to find the 3 that matter.
Notice what these have in common. Each one is a path the AI walks the same way, with your data, toward a finished thing. That is the engineering distinction Anthropic draws between a workflow and an agent in Building effective agents: a workflow follows predefined steps, an agent decides its own. Most of what saves you time is the boring workflow, not the clever agent.
Here is the part I will stand behind. The first time a skill writes something you would have written yourself, the whole relationship flips.
“An AI that knows you is impressive at a dinner party. An AI that does for you is the one you keep.”
From "an AI that knows me" to "an AI that does for me".
This is the line where the tool stops being a search engine.
I will give you my own, because abstract is cheap.
My weekly review used to be a Sunday evening I avoided. Open a blank page, try to remember the week, half-do it, feel guilty. Now a skill drafts it. This week it wrote the whole review against my actual calendar and call notes. I read it, fixed two lines, done. First time in a long while I did not start from an empty page.
Two changed lines. That is the whole story. The work went from an hour I dodged to five minutes I finished.
That is the shift the skills layer buys you. An AI that knows you is impressive at a dinner party. An AI that does for you is the one you keep. The first is a smarter search box. The second hands you finished drafts while you drink your coffee.
It only works on top of the layers under it. A skill with no memory is a generic prompt that writes a weekly review for nobody in particular. A skill with the wrong constitution drafts in the wrong voice and the wrong priorities. Constitution sets the rules, memory holds the context, skills do the work, and the layer above them, your tools, lets those skills reach out and touch your calendar and inbox directly.
You do not need all of it on day one. You need one skill that removes one recurring hour from your week. Mine was the Sunday review. Yours might be the proposal, or the inbox, or the Monday plan you never write.
A model is cheap. Context is expensive. And context that just sits in memory is a library nobody reads.
If you want an AI that does the work instead of only remembering it, that is what the skills layer is for. It is also the layer most people never reach, because they stop at storage.
Two ways in. If you have not built the memory underneath yet, start there: what a second brain actually is. If the rules come first for you, read the constitution, the first file. Then join the waiting list and I will walk you through your first skill.
I am still building mine. That is the point. You build one, use it, add the next.
- Anthropic, Building effective agents (2024): the workflow vs. agent distinction (a workflow follows predefined steps; an agent directs its own).
- Anthropic, Agent Skills (2025): composable folders of instructions and scripts a model loads on demand to run a repeatable task without relearning it each session.