Cheatsheet
Skillor agent?
A skill is a play: written down, it shows how something is done, and anyone can run it. An agent is a player: it has a shirt number, access to the pitch and a place in the squad — which means someone has to manage, watch and pay for it. Most tasks need a play. Every agent that should have been a skill is tomorrow's governance debt.
The two building blocks
Skill
Knowledge + instructionsThe play. A SKILL.md file with the steps and examples. It doesn't run by itself — no identity, no licence, no contract. Written in minutes, and it works in any team.
„Drop the folder into OneDrive under /Documents/Cowork/skills/. Up to 50 skills per person, one SKILL.md each. Cowork picks them up automatically at the start of a conversation."
Agent
Process + executionThe player. An actor with its own shirt number, access to the pitch and autonomy. It needs a squad place, supervision, a budget and a coach — every agent is a governance object.
„Lives in Copilot Studio, in Copilot and Teams, in Foundry Agent Service — and is governed through Agent 365."
Seven questions before you build
Who acts?
A play doesn't run on its own. Skill: you or an existing agent runs it. Agent: the task needs its own player on the pitch.
Identity and access?
Who gets a shirt number? Skill: none needed, it uses your context. Agent: its own identity and its own access to the data.
Autonomy?
Board or ball? Skill: shows the runs, doesn't play. Agent: keeps playing when you're not watching.
Effort?
Write a play or sign a player. Skill: a Markdown file and a clear description — and the description is what triggers it, so write it carefully. Agent: a new signing, with everything that follows.
Governance?
A collection or a coach? Skill: file versioning is enough. Agent: it needs supervision — that's what Agent 365 is for.
Reuse?
A good play works in any team. Skill: portable across platforms and agents, because the format is an open standard. Agent: tied to its own platform.
What does a wrong call cost?
This is the asymmetry. Skill: you delete a file. Agent: someone has to decommission an identity, its access and its budget — and until they do, it stays on the payroll.
The play first. Then the players.
Sign a player only when the task genuinely needs one. Every agent that should have been a skill is tomorrow's governance debt — like keeping someone in the squad whose only job is to read out the tactics board.
FAQ
How do I know whether a task needs an agent?
Ask whether something has to happen while you're not there. If the answer is no, a skill is enough. Autonomy is the dividing line, not complexity — a skill can describe a very complex procedure.
Where do custom skills live?
In your OneDrive under /Documents/Cowork/skills/, one subfolder per skill with a SKILL.md inside. Cowork finds them automatically at the start of each conversation. No admin, no package, no registration.
Why doesn't my skill trigger?
Almost always the description. Cowork matches the conversation against it, so it has to say plainly when the skill applies. Write it in the third person with concrete conditions, not as a title.
Can I move a skill to another tool?
Yes, that's the point of the open format. The same SKILL.md runs in other agent tools, provided the folder name matches the name in the file.
What does Agent 365 have to do with this?
It's the layer where agents are observed, governed and secured — the squad administration. Skills don't need it, which is precisely why they're cheaper to own.
Last updated: 6 August 2026
Every Monday at 7:30 CET, one tool you can put to work the same day.
Subscribe to the newsletter