Skills

Skills are versioned, reusable instruction sets that turn one good result into a permanent capability. Any agent can be assigned a skill.

Skills Registry

Skill Registry

The Skills page lists all skills in the current Workspace.

Skill List

Each row shows:

Name — Click to open Skill Detail
Category — e.g., Code Generation, Analysis, Documentation
Status — Draft / Published / Archived
Version — Current live version

Creating a Skill

Click New Skill → choose mode:

Mode	Description
Create Manually	Fill in name, category, description, instructions
Import from GitHub	Paste a GitHub URL (e.g., `https://github.com/anthropics/skills/tree/main/skills/pdf`) — auto-fetches SKILL.md
Import .ao	(Workspace Administrators) Choose a `.ao` skills package exported from this or another Workspace. Every skill arrives as a draft — review and publish it yourself. A name that already exists becomes `Name (2)`; existing skills are never changed

Workspace Administrators can download skills as a .ao file: Export all (Skills page) saves every skill in the Workspace, and the Export button in a skill’s header saves just that one. The file contains every version, the scripts and the uploaded files, so it can be imported into another Workspace or AO instance.

Skill Detail

Click a skill to open the detail view with tabs:

Skill Detail Instructions

Instructions Tab

The currently published version’s instructions. Read-only for published skills; editable for drafts.

Schemas Tab

Input/output JSON schemas for the skill (used for validation).

Scripts Tab

Executable tool scripts (e.g., hello_skill.py). Each script has:

Name
Language (Python, Shell, etc.)
Content editor

Files Tab

Additional files bundled with the skill (templates, configs, etc.).

Agent View Tab

Read-only preview of how the skill is materialized at runtime for agents:

On-disk layout under .opencode/skills/<slug>/
Scripts at scripts/<name>
Files at root
On-demand loading — agent loads full skill body only when needed via skill tool
Hard per-agent isolation — only agents assigned the skill can access it

Agent View

Versions Tab

Every edit creates a Draft version — never silently overwrites published.

Publish — Promote current draft to live (pins exact version)
Unpublish — Revert to draft
Archive — Hide from registry
Live badge — Shows (live) on the version agents actually receive

Native Skill Delivery

Skills are delivered at runtime as native opencode skills — not injected into prompts.

How It Works

Agent is assigned a skill
At run start, daemon materializes skill to <workdir>/.opencode/skills/<slug>/
opencode’s project-scoped discovery exposes only that run’s skills
Agent loads skill on-demand via skill tool (progressive disclosure)

Benefits

Hard isolation — Agent A’s skills invisible to Agent B
Version pinning — Published version fixed until re-published
Token efficiency — Only name/description in context; full body loads on demand
Script execution — Skill scripts run as native opencode tools

Skill Assignment

Agent Detail → Skills tab → Add Skill
Search/select from Workspace registry
Toggle Enabled per-assignment (can assign but disable)

Publishing Workflow

Draft (editing) → Publish → Live (agents receive) → Edit → New Draft → Publish → New Live

Only one live version at a time
Publishing pins exact version (agents get consistent behavior)
Draft changes don’t affect live until published

Recent Updates

Native Skill Delivery — Skills materialized as opencode skills per run, hard per-agent isolation, on-demand loading
Agent View Tab — Read-only preview of runtime materialization layout
GitHub Import — One-click import from GitHub repos with SKILL.md
Version Pinning — Publish pins exact version; draft changes don’t affect live