Step 2

Agents and skills

Skill or agent?

SkillAgent
What it isA procedureA colleague with a role
WhenA repeated task with a clear resultAn area that needs judgment and its own rules
ExamplePrepare an offerPetra, the sales assistant
Scheduled tasksNoYes

It is best to start with a skill. An agent makes sense once you need someone with their own area, tone and responsibility.

Create a skill or an agent

  1. In the brain, open the Skills or Agents tab.
  2. Click New skill or Hire an agent and fill it in like a job description. For scheduled tasks, just choose how often and at what time.
  3. Click Implement into the brain. A ready brief for the AI is copied to the clipboard.
  4. Open the brain in Claude Code or Codex, paste the brief (⌘V / Ctrl+V) and send it. The AI creates the skill or agent and saves it.

Each item shows its state: Draft only, Handed to the AI, In the brain.

System skills

Every brain has a set of commands from onlybrain that every role sees:

CommandWhat it does
/brain-helpwhat the brain can do: structure, skills, agents, connections, storage and what my role may do
/remembersaves information from the conversation in the right place in the brain
/findfinds the answer in the storage documents too and cites the exact page
/whats-newwhat changed in the brain
/tidy-upfinds duplicates, broken links, files without a description and outdated information
/describe-filesdescribes the storage files that have no description
/new-skillcreates a new skill or agent the right way for both Claude and Codex

They cannot be edited, onlybrain keeps them up to date. In the Skills tab, the manager can rename them and choose their language. These are the English names; in a Czech brain the commands are Czech (/pomoc, /zapamatuj, /najdi …). The help command is /brain-help because Claude Code already has /help.

Which of the brain's own skills and agents a role gets is set in the role (Sharing → role), by ticking them by name.

Where they live in the brain

  • Skill: skills/<name>/SKILL.md. Claude Code and Codex find it through short pointers in .claude/skills/ and .agents/skills/.
  • Agent: agents/<name>/CLAUDE.md with the instructions and a short subagent agents/_claude/<name>.md (onlybrain keeps the copy in .claude/agents/).

Both can be opened and edited in Files and editing. The pointers for the tools update by themselves when you save in the panel. It works the same on Windows, Mac and Linux, the brain uses no symlinks.

How to improve them

  • Add an example of a good result. A concrete sample works better than a long description.
  • Write down what the skill or agent must not do: send an e-mail or change a price without confirmation.
  • After a few uses, ask: "What kept repeating while using this skill, and what should be more precise?"