What a skill is
A skill is a set of written instructions, saved as a file, that an agent picks up when the work calls for it. It is how you hand an agent a way of doing something — your way — instead of explaining it again every time.
Each skill describes what it is for. When you ask an agent to do something, the agent matches the request against the skills available to it and loads the ones that seem relevant. You do not usually invoke a skill by name; it is pulled in because it fits.
Skills belong to the harness, not to Buzz. Buzz has no skill library of its own — the agent's harness holds them, which is why a skill has to be installed where that harness runs. See Model vs. harness.
Why it matters to you
Skills are the difference between an agent that is generally capable and one that works the way your team works. A model already knows how to write a summary; it does not know that your summaries lead with the decision, name the source, and never state a finding as validated before a specialist has seen it. A skill is where that lives.
This is usually the whole reason an agent is worth having in a channel. Without your skills, you have a competent stranger. With them, you have something closer to a colleague who has read your handbook.
Three distinctions are worth keeping straight, because they are three different places an agent's behaviour comes from:
| What it is | Who wrote it | |
|---|---|---|
| Instructions | The agent's standing role — what it does and does not do, always | You, when you set the agent up |
| Skill | Know-how for a particular kind of work, loaded when relevant | You or your team, as a reusable file |
| Memory | What this agent has retained from working with you | Nobody — it accumulated |
A skill is instructions, not ability. It tells an agent how you want something done. It does not give the agent access to a system it could not already reach — that is a separate decision. See Grant an agent access to real systems.
How to apply it
Write down the things you explain more than twice. That is the practical test for what deserves to be a skill.
Install the skills you want Buzz agents to use at the global level for your harness, not scoped to one project. A project-scoped skill is invisible to a Buzz agent. See Make skills available to your agents.
Be deliberate about what you install globally, because global means every agent on that machine — including agents in channels you would not have chosen for it.
Check rather than assume. Ask the agent what it has available, or give it a task only the skill can complete and look at the shape of the output.
Treat a skill from outside your team as something to read before you trust it. A skill is instructions going straight into an agent's head, so its author matters as much as its content.
Skills that arrive over the relay
Skills can also be published to a relay so a team can share them. Since v0.5.18, Buzz instructs agents not to fetch or follow a relay-published skill unless a person asks for that specific skill by name — and, even then, to treat its content as untrusted rather than as an instruction that outranks its own. Skills bundled with the harness or installed locally by you are unaffected. If you want a shared skill used, name it. Last checked: Buzz v0.5.20, 2026-09-02.
If you ignore this
The failure is silent, which is what makes it expensive. A missing skill does not produce an error and the agent does not say it lacks anything. It produces a plausible, competent answer that quietly is not how your team does it — and plausible-but-wrong is harder to catch than obviously broken.
People usually discover this the wrong way round: the output looks fine, gets used, and the gap surfaces later when someone who knows the method reads it properly.
Examples
A procedure in a folder on the shelf. A capable new colleague can do the work without it and will do it their way. Hand them the folder and they do it your way. Nobody has to stand over them.
A reviewer agent that checks a draft against your house style — the terms you have retired, the structure you use, the claims you never make without a source — because the skill saying so is installed and loads whenever a review is asked for.
Assuming an agent in Buzz can see the skills you use daily in your own tools. If they are scoped to a project rather than installed globally, it cannot, and nothing will tell you.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026