Skills
On this page
A skill is a written procedure Adverity Atlas can follow, "how we do X", packaged once and reused in chat. Instead of re-explaining your weekly budget review or your naming conventions in every conversation, you write them down as a skill and the assistant applies them when they fit. Skills are enabled for every organization.
What a skill contains
-
Instructions What to do, step by step, and what the answer should look like
-
Reference files Optional text material the assistant reads only when a step needs it, such as a CSV of benchmarks or a SQL snippet
-
Tools Optionally, the tools the skill needs. When a skill names tools, the assistant is limited to those built-in tools while it is active. Tools from connected MCP servers stay available.
mcp:entries are accepted but not validated and do not narrow MCP tools yet. A tool that is not available for your organization (feature off, or automation-only) is flagged on the skill's page and reported to the assistant, which tells you instead of guessing
A skill is a single SKILL.md file: a short header (name, description, tools) followed by the instructions in Markdown. Skills cannot contain scripts.
Adverity skills and organization skills
-
Adverity skills Maintained by Adverity and available to every organization. They cannot be edited, but you can fork one
-
Organization skills Written by your team in Settings > Skills, visible only to your organization
Each skill has a reference that tells the two apart: adverity/<name> for an Adverity skill and org/<name> for one of yours. An organization skill never replaces an Adverity skill with the same name, they sit side by side.
Using skills in chat
The assistant knows which skills are available and loads one on its own when your request matches its description. You can also choose one yourself:
-
Type
/at the start of the message box. A menu lists the skills you can use. -
Pick one. It appears as a pill above the message box.
-
Write your request and send. The assistant follows that skill for the conversation.
Typing a full reference such as /org/weekly-review check last week and sending attaches the skill and then sends the text after it as your message.
When the assistant loads a skill by itself, its tool list narrows from the turn after, not within the turn that loaded it. The tools stay those of the version the conversation started with, even if a newer version is published meanwhile.
To stop using a skill, click the x on its pill. Picking a different skill replaces the current one. If a skill is disabled or retired after you attached it, chat carries on with its normal tools and the assistant is told the skill is no longer available, so it stops following it. The pill stays until you remove it.
Skills for automations
Adverity's automation skills, and any skill whose category is automation, appear in Settings > Skills marked Automation. They are written for automations but you can also run one in chat from the / menu, for a one-off check. Each skill card and each version shows how many of your automations use it.
Writing your own skills
Go to Settings > Skills. The list shows each skill's reference, scope, version and status, and can be filtered by scope and status.
-
Create or fork. Start from scratch, or open an Adverity skill and click Fork to get an editable copy. A fork shows an upstream updated marker when Adverity publishes a newer version, with a diff you can read before editing your copy.
-
Edit. You fill in a form: description, category, inputs and tools, with the instructions in a separate box and a live preview beside it. Behind the form the skill is still stored as a
SKILL.mdfile. Upload reference files underreferences/(.md,.csv,.sqlor.json, up to 256 KiB each, at most 20 files and 1 MiB in total). Edits only change your draft. Discard draft drops the draft and keeps the published version. On a skill that was never published it reads Delete skill and removes the skill for good, which frees its name. -
Try it. Your draft appears in your own
/menu, tagged as a draft, so you can test it in chat before anyone else uses it there. -
Publish. Review the diff, add a change note, and publish. Published versions never change afterwards, publishing again creates a new version.
-
Review history. The detail page lists every version and shows a diff between any two.
-
Disable or retire. Disable a skill for the whole organization or for one workspace to hide it from chat, or retire it to take it out of chat. A Knowledge Manager or Admin can undo a retirement with Restore on the skill's page, which brings back its last published version. A retired skill's name stays reserved.
Limits: the description is at most 1,024 characters (it is what the assistant matches your request against), and the instructions at most 500 lines and 24,000 characters.
You can also ask the assistant to help. The Adverity skill-authoring skill teaches it the format and what makes a good skill.
Write a skill in chat
Ask the assistant to write or change a skill. It interviews you, shows the full SKILL.md, and then offers to save it. A card shows the description and what changes (a diff when the skill already exists); nothing is saved until you approve. The result is a draft you can try in your / menu, and a Knowledge Manager or Admin publishes it from Settings > Skills. Editors can save drafts this way.
Who can do what
Scroll horizontally to see the full table.
Role | Can |
|---|---|
Viewer | Use skills in chat and read them in Settings |
Editor | Also create skills, fork, and edit drafts |
Knowledge Manager | Also publish, retire, restore, and disable skills |
Admin | Same as Knowledge Manager, plus everything else admins manage |
In chat, a draft belongs to whoever saved it last: only they see it in the / menu and the assistant's skill list, and everyone else gets a skill only once it is published. When someone else saves the draft, it becomes theirs and you go back to the published version. In Settings > Skills, only people who can create skills (Editors and Admins) see drafts, and a skill that was never published is hidden from Viewers.