Skip to content

Skills

A skill is a folder with a SKILL.md, in the form Claude Code reads: a name and a description at the top, instructions below, and files beside it. A role that has a skill is told its instructions and can use its files.

Skills are on the Tools page, tab Skills: "KOLONİ's own store: a copy no agent writes."

Adding a skill

Press Add → A skill · from a folder or GitHub. Then:

  • Skill folder: type the folder's full path and press Read the folder. Or Choose folder… (at most 5 MiB) or Upload .zip… (at most 25 MiB) from your browser.
  • From GitHub: a public repository, or a folder in it, such as https://github.com/owner/repository/tree/main/folder. Press Read from GitHub. First only the list of files, the SKILL.md, the licence and the scripts are read; the rest is downloaded when you add it.

KOLONİ keeps its own copy. Secrets and links in the folder are not copied.

Before you add it, a card shows what the skill holds: its licence (or that it is unknown), its files, which files run as programs, what was not copied, and a look at its scripts. Show SKILL.md shows its text. Press Add skill, or Replace skill if a skill of that name is kept already.

Example: skills from superpowers

superpowers is a public set of skills under the MIT licence; each is a folder under skills/. To add its test-driven-development skill for the Coder:

  1. Press Add → A skill · from a folder or GitHub.
  2. In From GitHub, type https://github.com/obra/superpowers/tree/main/skills/test-driven-development and press Read from GitHub.
  3. Read its SKILL.md on the card (Show SKILL.md), then press Add skill.
  4. Open the skill's row and tick the Coder under Roles that may use it.

Some of them may fit these built-in roles:

Skill May fit
test-driven-development Coder
systematic-debugging Debugger
verification-before-completion Coder, Tester
receiving-code-review Coder
requesting-code-review Coder; the Reviewer may use its template

Don't give them to the Orchestrator: its own way of working is KOLONİ's, and a workflow skill there (brainstorming, writing-plans, using-superpowers) makes it plan and decide in your place.

Several are written for Claude Code's own way of working: they name its tools, worktrees, sub-agents and a browser companion. These fit KOLONİ's roles poorly: brainstorming, subagent-driven-development, using-git-worktrees, executing-plans, dispatching-parallel-agents and using-superpowers. A skill's row says when a role lacks a tool its allowed-tools line lists; tools named only in its text are not checked. brainstorming and systematic-debugging hold scripts: on the card, before Add skill, untick Run the scripts of … without asking unless you have read them.

A skill steers every agent that has it, so give each one only to the roles that need it.

Scripts

Files ending in .py, .js, .mjs, .cjs, .sh or .ps1 count as scripts. Run the scripts of … without asking is on by default. "Saying yes to this is agreeing to run them. Any other command is asked about as before, and once a file of the skill has changed since it was added its scripts are asked about too."

A skill's scripts need the role to be able to run commands, and they do not run in projects whose commands run in Docker.

Giving a skill to roles

Open the skill's row and tick Roles that may use it. If a role lacks a group of tools the skill needs, the row says so and offers to give it.

A skill's text is what the agents that have it are told, so a skill from elsewhere can steer them. Read it before you give it to a role.

When an agent proposes a skill

An agent can propose a skill. The Add this skill? card shows the same things as above, and which roles it would be given to. Press Add skill or Reject.

How a skill did

Open a skill's row: Did lately shows the last 30 days, one row per provider, model and role that had the skill. Its bars show Taken in, Cost or Steps, once there are 5 jobs or more. Each row says how many of those agents read the skill whole ("read in 2 of 6"), or "given, not read", in grey: then its numbers say little of the skill. An agent that cannot read a skill itself (it has no use_skill) is given its skills in full in its brief instead. The row counts those too, as "given in full in its brief in 3 of 6", not as "given, not read".

Switching off and deleting

Each skill has an on / off switch. Delete skill removes it: "Every role that has it loses it, and KOLONİ's copy of it is removed."