Reusable instruction sets for recurring tasks. Type / to open the catalog, attach a skill as a chip, then add your task.
Staged changes3 files
src/auth/session.ts+18−4
src/auth/rotate.ts+31
tests/session.test.ts+22
Skills
Prepare a skill instruction in the ML.ai composer.
Teams with repeated workflows like commits, subagent creation, or project-specific automation who want consistent, reusable instructions.
Project skills live in the repo, so everyone runs the same steps for commits and reviews.
Save the instruction once and call it from the / list instead of retyping it.
Package project-specific steps as a skill the whole team can pick up.
A skill is a file with frontmatter and instructions. Type / to pick one, it attaches as a chip, and the agent follows it for that task.
name: release-notes
release-notes/name: release-notes
SKILL.mddescription: Draft notes from git log
example.md---
A skill is one file with frontmatter
The frontmatter names the skill and says what it is for, and the body holds the steps. Keep it in the project for your team or make it global for every repo.
src
/commit
/create-subagent
/release-notes
Write a commit message from staged diff
Scaffold a new subagent definition
project: draft notes from git log
Defaults, project and global in one list
Type / in the composer to open the catalog. The five default skills sit in the same list as your project and global ones, so there is one place to look.
Same skill, same format every time
A skill carries the steps you would otherwise retype. Attach it and the notes follow the same sections, order and commit links on every release.
Use this message for the 3 staged files?
Skill attached: commit
Chip on the task, approval before it runs
The chip sits in the composer next to what you type. The agent reads the skill first, then your task, and any command it runs still asks for approval.
Pasted again from last week, slightly different this time
/commitDefault
/create-subagentDefault
/release-notesProject
You paste the same instructions into each task, they drift from the last version, and every teammate ends up with a slightly different way of writing commits or release notes.
Type /, pick the skill and it attaches as a chip. Default, project and global skills all show in one list, and everyone on the repo picks up the same instructions.
Writing your own skill takes one file. Here is the path from a task you keep repeating to a chip anyone can attach.
· Recent promptsWrite release notes for this weekMon
Write release notes for the hotfixWed
Write release notes for v2.4Fri
Same instructions each time, a good fit for a skill
· New skill · Save skillWhere should release-notes live?
ProjectLives in this repo, so everyone on the team gets it
GlobalFollows you into every repo you open
/commitDefault
/create-subagentDefault
/release-notesProject
· Recent promptsWrite release notes for this weekMon
Write release notes for the hotfixWed
Write release notes for v2.4Fri
Same instructions each time, a good fit for a skill
What trips people up when they start writing skills.
What goes wrong
How handles it
Pasted instructions change a little each time, so the same task comes back in a different shape from one week to the next.
How handles it
The instructions live in one skill file. Attach it with / and every run starts from the same text.
A useful prompt saved in one person’s notes does nothing for the rest of the people working in the repo.
How handles it
Project skills sit in the same / catalog as the defaults, so anyone working in the repo sees them as soon as they type /.
Long skill files try to cover every case and end up hard to follow, for the agent and for you.
How handles it
Keep each skill to one job with a clear description in its frontmatter, and split bigger workflows into separate skills.
SettingsMCP servers
Connect local or remote tools that ML.ai can use while working.
Add server
Local server started by a command
Connect local or remote tool servers so a skill’s steps can reach your databases and APIs.
SettingsAgents
Pick a subagent and it stays selected until you swap it. Only general carries out changes end to end.
@explore · Where is the refresh token rotated?
› Searched refreshToken · 14 matches
› Read src/auth/session.ts, src/auth/rotate.ts
› Answered · 0 files changed
@exploreAsk ML.ai... @ to delegate
Searched getUser9 matches
Edited 4 files+12 −12
Ran pnpm test48 passed
Inside the project only · back to you after one turn
Edits and commands, inside the project
The four subagents, what each one can do, and how to pick one with @ alongside a skill chip.
Write a file with frontmatter that gives the skill a name and a description, then save it as a project skill or a global one. It shows up in the / catalog next to the five defaults.
Make it a project skill when the steps belong to one repo: it lives with the project, so everyone working there gets it. Make it global when it is your own workflow and should follow you into every repo you open.
A name and a description that says what the skill is for; the steps go in the body. A skill with malformed frontmatter is skipped by the engine rather than shown as broken, so check it first if a skill does not appear.
No. A skill adds instructions to the task, not permissions: any command the agent runs while following it still asks for approval as usual.
Five: commit writes a Conventional Commits message from your current diff, create-subagent composes a custom subagent from a plain-language description, pixis-proof explains what a run’s verification evidence establishes, project-pdf generates a PDF overview of the project, and site-redesign-ui-ux designs, audits or implements web interfaces.
Write a skill once, attach it with / whenever you need it, and give the whole team the same instructions.
