Skills
A skill is a set of instructions for one task, such as “review my week” or “model in Blender”. You keep each skill once, in your Skills collection, and every agent you use reads the same copy.
The format
Every skill uses the open Agent Skills format. Claude Code, Codex, OpenCode, Cursor and
others read it. A skill is a folder with a SKILL.md file (a few fields on top, then the
instructions) and any scripts or files it needs:
Collections/Skills/ blender-modeling/ SKILL.md scripts/export_glb.py weekly-review/ SKILL.mdIn Kaizendex, a skill is an entry. Its top fields are properties (name, description, which agents, which projects, version, enabled). Its instructions are the entry’s text. Its scripts are attached files. The folder above is a copy the app writes so other agents can read it.
Use one skill in many projects
On the desktop app, a skill’s Use in projects list shows your coding projects. Tick a
project, and the app puts a link to the skill folder into that project’s
.claude/skills/ and .agents/skills/. A link is a pointer, not a copy, so there is one
skill to keep up to date.
- Managed copy: for a tool that cannot follow links, the app keeps a real copy in step.
- Global: one switch puts a skill in
~/.claude/skillsand~/.agents/skillsfor every project. The app warns you: every global skill costs the AI some reading on every turn. - Edits come back: if Claude Code edits a linked skill inside a project, it edits your one copy. The change is logged, synced and undoable.
Slash commands
Type / and a skill’s name in a chat to call it, for example /weekly-review. Slash commands
are skills, not a separate list.
Skills that get better
The AI improves a skill from how you use it. Example: the AI made a Weekly review skill for you.
You run /weekly-review and say “put wins first”. The AI edits the skill by itself, and the chat
shows “Updated skill: Weekly review · See · Undo”. Next week, wins come first.
What counts as a lesson:
- you correct the result;
- you redo or edit what the AI made;
- an automation failed, then worked after a fix;
- you accept a result as it was.
The AI’s own grade of its work never counts alone.
Small edits, never rewrites. Every change is in the skill’s history, with before, after and Undo.
| Skill | What the AI does |
|---|---|
| A skill the AI made | improves it by itself, with Undo; it can also make a new one: “New skill: Weekly review · See · Undo” |
| A skill you wrote, pinned, or took from the Library | proposes a change, shown as a diff (what changed, line by line), and waits for you |
| A script in any skill | never runs without your approval |
Limits:
- At most 50 active skills made by the AI (a setting).
- A skill’s text is up to 15,000 characters.
- A skill the AI made and nobody used for 30 days is proposed for the archive, in the weekly tidy-up. Pinned skills never are. An archived skill comes back in one tap.
Did it help? Each skill counts its uses, and the corrections after each use. A skill that keeps getting corrected is flagged in the weekly tidy-up.
You can switch this off: untick Learn and improve skills by itself in Settings → AI → Memory. See How your AI learns.
Skills from others
You can import a skill from the Library or from GitHub.
- A skill never grants powers (by default). It is only instructions. It gives the AI no new tools and no new permissions, whatever its text says. See AI safety settings.
- Your copy stays on the version you took. An update shows a diff per file (what changed, line by line) and waits for you.
- The app scans each skill for hidden instructions aimed at the AI.
- Scripts arrive switched off. Until you approve them, they stay outside every linked folder, so a coding agent cannot run them by accident.
You can publish your own skills. You publish each one on purpose, one at a time, after you see its full text. A skill never goes out inside something else, such as a hub or an agent. See Share agents and skills.