Skills
Give any agent reusable, install-once capabilities — attach from Agent Home, manage from Settings
Overview
A skill is a folder with a SKILL.md file that teaches an agent how to do something — drive a CLI, follow a house convention, use a tool. Agentastic's Skills feature gives you one place to collect skills and one gesture to hand them to any agent, instead of copying files into each agent's config by hand.
Two things are kept separate:
- Library — your personal collection of skills, managed in Settings → Skills. This is the source of truth and lives outside any one project.
- Delivery — when you attach a skill and launch, Agentastic copies it into that worktree, in the format the target agent expects. Nothing is written to global agent config, and parallel launches with different skills never race each other.
Skills work across Claude Code, Codex, Droid, and Amp, each via its own on-disk convention (see Supported Agents below).
Attaching a Skill in Agent Home
There are two ways to attach a skill to a launch, both backed by the same library:
Type /
Type / at the start of the Agent Home prompt (or the start of a new line) to open a picker: your installed skills on top, matching results from the Skillopedia catalog below. Keep typing to filter, use the arrow keys and Enter to select, or click a row.
- Picking an installed skill attaches it immediately.
- Picking a Skillopedia result installs it into your library first, then attaches it — one step either way.
The + menu
Click + below the prompt and choose Attach skill… for the same picker in a popover with its own search field. Useful when you don't want to type / inline, or want to attach a skill before writing the rest of your prompt.
Attached skills appear as removable chips above the prompt — click the ✕ on a chip to detach it before sending.
What the Agent Receives
On launch, Agentastic copies each attached skill into the new worktree using the target agent's native skills directory (for example .claude/skills/<name>/ for Claude Code), and appends a short note to the prompt so the agent knows it was chosen on purpose:
The user selected the `superpowers` skill for this task. Invoke it with `/superpowers`.
Agents that auto-invoke skills by description (Claude Code, Amp) get a lighter note telling them the skill is available to use; agents that expect an explicit invocation get the exact command.
Delivery is worktree-scoped and happens fresh on every launch — updating a skill in your library takes effect the next time you attach it. In local mode (the worktree is your actual repo), the per-agent skills directory is added to .git/info/exclude automatically, so delivered skills never show up as untracked changes.
Managing Your Library (Settings → Skills)
Open Settings → Skills to install, update, or remove skills.
Install from GitHub
Enter a public repository as owner/name (for example anthropics/skills) and click Install. Agentastic finds the first SKILL.md in the repo and installs that folder. This is unauthenticated, so it's subject to GitHub's public rate limit — fine for occasional manual installs.
Install from Skillopedia
Search the built-in Skill Marketplace field to browse Skillopedia's catalog and install a result with one click. Skillopedia tracks a content hash for each skill, which is what powers Update to Latest.
Update and remove
Every row in Installed Skills has a menu with Open in Editor (opens the skill's SKILL.md directly), Update to Latest (for GitHub- and Skillopedia-sourced skills), and Remove.
Bundled Skills
Agentastic ships with five skills pre-installed, so there's something useful on day one:
| Skill | What it does |
|---|---|
browser-use | Drive Agentastic's built-in browser via dev browser commands — navigate, inspect, fill forms, verify UI state. See Browser Skill for Agents. |
dev-cli | Drive a running Agentastic instance from outside its terminals using the system-wide dev CLI. |
skillopedia-discovery | Search and install more skills from Skillopedia on the fly. |
superpowers | obra/superpowers — a general skill-discipline framework that establishes how to find and use skills before responding. |
gstack | A headless-browser QA toolkit for testing and dogfooding a site — open pages, interact, verify, capture screenshots as bug evidence. |
Supported Agents
Each agent reads project skills from its own conventional directory:
| Agent | Worktree skills directory |
|---|---|
| Claude Code | .claude/skills/<name>/ |
| Codex | .agents/skills/<name>/ |
| Droid | .factory/skills/<name>/ |
| Amp | .agents/skills/<name>/ |
Other agents can still read a skill manually — the delivered folder is a plain SKILL.md-rooted directory, and nothing about it is Agentastic-specific.
Writing Your Own Skill
A skill only needs a SKILL.md with a name and description in its YAML frontmatter:
--- name: my-skill description: Use when the user asks to do the specific thing this skill teaches. --- # My Skill Instructions for the agent go here.
description is what auto-invoking agents use to decide when to pull the skill in, so make it specific about when the skill applies. Anything else in the frontmatter (for example Claude Code's allowed-tools) is preserved and delivered as-is — Agentastic only needs name and description to list the skill in pickers.
To share a skill with your team, install it from a GitHub repo everyone has access to; to share it with anyone, publish it to Skillopedia.
Tips
- Attach before you type — use the + menu's Attach skill… when you want to pick a skill before drafting the rest of the prompt.
- One library, every agent — install a skill once; it's available to whichever agent you launch next, translated into that agent's own convention.
- Keep the library curated — remove skills you don't use so the
/picker and Skillopedia results stay easy to scan.