UniloomDocs
Use Uniloom

Adding items

Press + in the Toolkit side panel to add an item. On a tab it opens that kind's form, such as New skill on Skills. From All projects, it first asks where the item goes: User or one project.

Where an item can live in several places, the form shows a card for each place with the agents that read it. Tick one or more. A line under the cards names every agent that will read it. Create writes the files; Cancel writes nothing.

Skill#

Give it a name (lower case letters, digits and dashes, such as review-pr), what it does and when an agent should use it, and where it goes.

  • With .uniloom/ in the project, Every agent: written once, published to each puts it in .uniloom/skills/, with a group and a blank file or a copy of another skill to start from. The other places are then already covered.
  • Without .uniloom/, pick the agents' folders, such as .claude/skills or .agents/skills. The skill is written once and the other folders link to it.

Who starts it: You or the AI, Only you or Only the AI. Claude Code reads this; the other agents let both start a skill.

Command, subagent, rule, output style#

Each takes a name, a few fields, and the agents' folders. The Writes box shows each file's header in its agent's own format before you press Create.

  • Command: a description, and an Argument hint for Claude Code.
  • Subagent: When to hand work to it, a Model (Same as the session, Fast or Strongest), and Tools it can use (All, Read only, or picked from Read files, Search, Edit files, Run commands and Web). Each agent gets its own model name, such as Claude Code's haiku or opus. OpenCode keeps the session's model.
  • Rule: When it applies: Always, When files match (patterns such as packages/server/**), When relevant or Only when mentioned. Claude Code reads a rule only always or by files, so the last two leave it out. A Codex rule is a command policy: its first words, a Decision (Allow, Ask first or Forbid) and Why.
  • Output style: a description and Keep coding instructions. On, the style adds to Claude Code's coding instructions; off, it replaces them.

Instructions, MCP server, hook, plugin#

  • Instructions: the instructions files the project does not have yet.
  • MCP server: the form opens on the apps Integrations can connect, such as Linear, Notion and Jira, each with its logo. Press one to add its official server to the agents ticked under Add it to. It is written as a URL with no keys: each agent signs in to the app in your browser the first time it uses it. An app already in the project shows a check and cannot be added again. An agent that cannot sign in on its own is left unticked, and a line under the list names it.

Other server opens the full form: a name, a Command or URL, environment variables (one NAME=value a line), and the agents. Each ticked agent shows the file it goes in. The variables are stored as plain text in files at the project's root, which are often committed, so keep real keys out of them.

  • Hook: the Agent (Claude Code, Gemini, Cursor or Qwen Code), an Event, an optional Matcher such as Bash|Edit, the Command, and a timeout (seconds or milliseconds, as the agent counts). A Claude Code hook can stay Only on this computer.
  • Plugin: its name as the agent names it (name@marketplace for Claude Code and Codex, an npm name for OpenCode), and the agents to Turn it on for. Installing it stays the agent's own command.

Last updated