Using Skillshub

Skillshub is a package manager for AI coding-agent skills. Think Homebrew, but for the SKILL.md files that agents like Claude Code, Codex, Cursor, OpenCode, Continue, etc. read for instructions. One install, many agents — skillshub link wires installed skills into every detected agent.

This skill teaches you (the agent) how to drive the CLI on the user's behalf.

When to reach for which command

Match the user's actual intent before picking a command. The most common mismatch is reaching for tap add when the user just wants one skill from a repo — skillshub add <url> is usually the faster path.

The full command surface lives in references/cli-reference.md — load it when the user asks about something not in the table above, or when you need exact flag names.

The mental model

Three locations matter. Confusing them is the source of most "why doesn't my agent see this skill?" questions:

GitHub tap repo  ──git clone──▶  ~/.skillshub/taps/owner/repo/   (cache)
                                          │
                                          │ install copies the skill
                                          ▼
                                ~/.skillshub/skills/owner/repo/skill/   (canonical)
                                          │
                                          │ link symlinks per-skill into…
                                          ▼
                                ~/.claude/skills/, ~/.codex/skills/, …  (agent dirs)

• Taps are git repos with SKILL.md files anywhere inside. Skillshub

walks the clone to discover them — no manifest required.

• Installed skills live in ~/.skillshub/skills/<owner>/<repo>/<skill>/.

This is the canonical copy.

• Linked skills are per-skill symlinks from each agent's skills directory

back to the canonical copy. This is why one install fans out to every agent.

• External skills are SKILL.md folders an agent already had (e.g. from a

marketplace). skillshub link discovers them and replicates them to the other agents so everyone stays in sync.

If the user reports "agent X doesn't see skill Y", check in this order:

1. Is the skill installed? skillshub list — look for it.
2. Is agent X detected? skillshub agents.
3. Has skillshub link been run since installing? Re-run it; symlinks are

not created automatically on install.

1. Run skillshub doctor for git/clone/orphan diagnostics.

Common workflows

Install one skill from a URL

The user pasted a GitHub URL pointing at a skill folder.

skillshub add https://github.com/owner/repo/tree/main/path/to/skill
skillshub link

Use add (not tap add) when they only want this one skill — it bypasses tap management entirely and creates a synthetic owner/repo namespace behind the scenes.

Subscribe to a tap and install everything

The user wants a whole collection.

skillshub tap add anthropics/skills --install
skillshub link

Without --install you get the registry but no installed skills; you'd then pick with skillshub install anthropics/skills/<name> or grab them all with skillshub tap install-all anthropics/skills.

Add a skill from a Gist

Gists are flat (no folders), so each file with valid SKILL.md frontmatter becomes its own skill under owner/gists/<skill-name>.

skillshub add https://gist.github.com/user/<gist_id>
skillshub link

Set GITHUB_TOKEN first to avoid rate limiting — gists hit the GitHub API, unlike regular taps which use git clone.

Sync after editing or installing anything

Anytime installed skills change — first install, update, uninstall, or after the user manually drops a skill into one agent's directory — re-run:

skillshub link

It's idempotent and cheap. Suggest it whenever the user reports sync-feels-off symptoms.

Update

skillshub tap update             # refresh all tap registries
skillshub update                 # update every installed skill
skillshub update owner/repo/skill   # one specific skill

Tap update and skill update are separate steps because the registry (what's available) and the install (what's on disk) are tracked independently.

Uninstall a tap but keep its skills

Useful when the user wants to stop tracking a registry without losing what's already linked into their agents.

skillshub tap remove owner/repo --keep-skills

Full purge

skillshub clean all              # interactive confirm
skillshub clean all --confirm    # for scripts/CI

This removes every skillshub-managed symlink from every agent dir and deletes ~/.skillshub/ entirely.

Writing or editing a SKILL.md

When the user asks you to author a new skill or fix one, follow this format. Only name is required; everything else is optional but worth filling in.

---
name: skill-name
description: One-sentence trigger — what the skill does AND when to use it.
allowed-tools: Read, Edit, Bash       # optional, comma-separated or YAML array
license: MIT                          # optional, SPDX identifier
metadata:                             # optional nested block
  author: my-org
  version: "1.0"
---

# Skill Name

Imperative instructions for the agent...

A few rules worth internalising:

• The description is the trigger. Skillshub doesn't drive triggering, but

Claude does — it picks skills based on the description. Make it specific about when to fire, not just what it does.

• One folder per skill. The folder name should match name.
• Optional sibling dirs: scripts/ for executables the agent can run,

references/ for deeper docs the agent can load on demand.

• Auto-discovery. Anywhere in a tap, any folder containing SKILL.md is

a skill. No manifest, no registration. This means the user can drop skills under skills/, experimental/, agents/, wherever — skillshub will find them.

Supported agents

Skillshub auto-detects these and links per-skill into each:

Claude (~/.claude), Codex (~/.codex), OpenCode (~/.opencode), Aider (~/.aider), Cursor (~/.cursor), Continue (~/.continue), Trae, Kimi, OpenClaw, ZeroClaw, Kiro (uses ~/.kiro/steering), Gemini, Copilot, Junie, Augment, Warp, Cline.

Run skillshub agents to see which are present on the user's machine.

Things to watch out for

• @commit no longer works for non-gist taps. Shallow clones can't check

out arbitrary commits. If the user pastes owner/repo/skill@<sha>, expect a hard error and suggest installing latest instead.

• Private repos need git credential helpers or SSH keys configured. A

GITHUB_TOKEN is not enough for tap operations — only for gists and star-list imports, which still hit the GitHub API.

• git is a hard runtime dependency. If skillshub doctor reports a git

problem, fix that first; nothing tap-related works without it.

• External skills are real. When the user manually drops a skill into one

agent's directory, skillshub link will replicate it to every other agent. This is intentional sync behaviour — flag it if it might surprise them.

Where to dig deeper

• references/cli-reference.md — every command and flag
• references/architecture.md — directory layout, data flow, key crates
• The skillshub repo itself: https://github.com/EYH0602/skillshub