# Skills & extensions Source: https://docs.ideharmony.com/customization/skills Skills are named, versioned, reusable workflows that agents can invoke: the difference between an agent that improvises every task and one that follows your playbook. Harmony ships with a built-in set, and the skills manager keeps them durable and inspectable. ## The built-in skills | Skill | What it encodes | |---|---| | `code-review` | A structured review pass: findings, severity, and resolution. | | `test-runner` | Run the suite, parse failures, report precisely. | | `doc-generator` | Produce documentation from code and context. | | `bug-triage` | Reproduce, isolate, and scope a reported bug. | | `refactor` | A disciplined refactor loop: plan, transform, verify. | Beyond the built-ins, Harmony ships reusable workflows for planning, debugging, reviewing, debating, documenting, and breaking work into tickets, alongside the 40+ agent presets in [Company mode](/docs/orchestration/company-mode). ## Managing skills From the TUI: ``` /skills list and manage installed skills /extensions list and manage extensions ``` Skills are stored durably by the skills manager service, with CRUD and versioning, so an agent improvement you make on Tuesday is still there on Friday, in every workspace. ## Extensions Extensions have an isolated lifecycle (install → enable → disable → uninstall) so a misbehaving extension can be switched off without surgery. Because Harmony's agents are assembled from the open Pi framework, Pi's whole extension ecosystem snaps in too: the workspace that runs your fleet doubles as the kit for building your own assistant, voice agent, or something nobody has named yet. ## Tools and safety Agent tools are gated by **Settings → Tools**: safe tools (inspect agents, prompt them) are on by default; dangerous tools (spin up an agent, open windows or terminals, spawn subagents) start off. Turn on dangerous tools deliberately, and let [approvals](/docs/concepts/permissions-and-approvals) catch anything that still deserves a signature. ## Writing your own A skill is a named workflow with a version: clear trigger conditions, the steps to follow, and the verification to run before declaring done. Start by watching how a built-in like `code-review` structures its pass, then write yours in the same shape. When it works, version it, and let every agent in the fleet use it.