Skills

Skills add reusable procedures to AI workers. The pieces people miss in order: (1) the built-in catalog is hard to keep in your head, (2) some built-ins are 'meta-skills' that let workers reshape Zemu itself, (3) the three custom-skill on-ramps each fit a different scenario, and (4) Workflow Skill is a different runtime mode from regular skills.

Overview

What the 19 official built-in skills cover, the meta-skill family (agent-builder / skill-manager / etc.), three custom-skill paths, and how Workflow Skills differ from regular ones.

Basics

The 19 official built-in skills

The dashboard Skills page lists Zemu's 19 official (built-in) skills, all marked with the "Official" badge. Grouped by job:

  • File ops: docx / pdf / pptx / xlsx
  • Image generation: gemini-image-generation / gpt-image-generation
  • Agent operations (meta-skills): agent-builder / agent-config / skill-manager / agent-browser / knowledge-manager
  • Search & research: people-search / company-search / web-scraper
  • Voice & comms: voice-call and others

Every worker can see these candidates, but none of them runs unless explicitly enabled on the worker. Filter the list by search / tag / creator.

Meta-skills let workers reshape Zemu itself

Four built-ins — agent-builder, agent-config, skill-manager, knowledge-manager — are meta-skills that change Zemu's own configuration.

  • agent-builder: create new workers (instructions, skills, integrations included)
  • agent-config: change settings on existing workers
  • skill-manager: create / inspect / edit custom skills
  • knowledge-manager: read and write the knowledge tree

A worker with these enabled can be asked in plain language to "create a new worker for X" or "edit this skill" and it will rewrite Zemu's setup. Conversely, workers without meta-skills cannot modify Zemu — useful as the default permission boundary for end-user-facing workers.

Three on-ramps for custom skills

The + Add menu in the top right offers three creation paths.

  • Build with Agent: ask Zemu in chat to skill-ify a procedure; a worker holding skill-manager writes the SKILL.md and workflow definition for you
  • Upload a skill: upload .skill / .zip / .tar.gz packages — best for moving skills between environments
  • Import from GitHub: paste a repo URL; a directory containing SKILL.md is pulled in as a skill

A common shape: prototype the skill via chat, then move ownership into a Git repo and import from GitHub for change history and reviews.

Workflow Skill — runs end-to-end from a trigger

Regular skills are called partially by a worker when it judges they are needed. A Workflow Skill is invoked whole: the predefined sequence runs from start to finish.

In the schedule-trigger creation dialog, switching the execution-method toggle to workflow_skill selects the Workflow Skill at trigger time and runs its steps in order — distinct from agent_prompt which hands a prompt to the worker.

Decision rule: does the work change every run, or is it always the same?

  • Steps are fixed → use a Workflow Skill from the trigger; updating the skill propagates to every trigger that calls it (easier maintenance)
  • Inputs vary → use agent_prompt, where the worker interprets the prompt at runtime

When multiple triggers should share the same logic (weekly + monthly reports, for example), keep the Workflow Skill single and pass the date range as a parameter from each trigger.

Checklist

  1. On each worker's Skills tab, can you list every enabled skill and explain why it is on?

  2. Is the boundary between workers with meta-skills (agent-builder etc.) and workers without them deliberate, not accidental?

  3. When automating a recurring task, do you have a rule for choosing Workflow Skill vs. agent prompt?