SKILL.md that starts with YAML
frontmatter, followed by Markdown instructions. Skills follow the
Agent Skills format, so skills written
for other agents work here too.
SKILL.md
name uses lowercase letters, digits and single hyphens, at most 64
characters, and matches the folder name. The description says what the skill
does and when to use it, in at most 1,024 characters.
How the agent uses skills
Each chat message the agent receives lists the available skills with their descriptions andSKILL.md paths. When a task matches a description, the agent
reads that SKILL.md and follows it, opening the skill’s other files only when
needed. A long reference document therefore costs nothing until it is used.
Unattended automation runs do not receive the list, because it would use up
their small, approved token budget.
The easiest way to add a skill is to ask the agent: “Create a skill for our
weekly status update”, or “Install the skill from
https://github.com/example/skills/tree/main/pdf”. The agent writes personal
skills to /home/r5d/.agents/skills/<name>/, and they are listed from the next
message.
Where skills live
When two skills have the same name, the first location in this table wins, so a
user can adapt a shared or built-in skill by copying it into their personal
skills.
Manage skills in the browser
Settings → Skills lists every skill the user’s agent can use. Opening one showsSKILL.md rendered as Markdown, with the frontmatter as labelled
metadata, beside the skill’s files. Users can:
- create a skill from a template, and delete a skill;
- edit any text file in an editor with syntax highlighting, and save with Ctrl/⌘+S;
- add files and folders, upload files or a whole folder, and rename, move, download or delete files;
- mark scripts executable;
- copy a shared or built-in skill into their personal skills.
SKILL.md, such as a missing description or a name that does not
match the folder, are shown above the file without preventing a save.
Personal skills are files in the user’s workspace, so the agent and the browser
always see the same skill. Browsing never starts a stopped workspace: the list
shows the personal skills last seen, and opening or changing one asks to start
the workspace. A save carries the version of the file it was based on; if the
agent changed the file in the meantime, the editor offers to reload it or
overwrite it, so neither change is lost silently.
Deleted personal skills are moved to /home/r5d/.r5dchat/trash/skills and
permanently removed after seven days. Ask the agent to restore one in time.
Shared skills
Administrators manage shared skills under Administration → Skills and in Settings → Skills. They are stored by the control plane, so they reach every user, including users whose workspace has never started.- A new shared skill is a draft: only administrators see it, and it is not copied to any workspace. Publish it when it is ready.
- The control plane copies the published shared skills into each workspace when the workspace is next used, and immediately into running workspaces after a change. The workspace copy is replaced as a whole on every update, so changes made there by an agent do not persist.
- An administrator can share one of their own skills, or a built-in skill, with everyone. The copy is published straight away unless they choose to keep it as a draft.
Skill files are served to the browser as downloads, never as pages: only
raster images display inline. Markdown is rendered with the same sanitizer as
agent responses. Treat skills from unknown sources like any other code, and
review their instructions and scripts before relying on them.