What does this skill do, and when should you use it?
Skills Manager is an open-source (MIT) desktop app that stores each skill once in a central hub (~/.skills-manager/skills) and symlinks it into 32 AI coding tools including Claude Code, Codex, Cursor, and Gemini CLI. skm is the companion command-line front-end that reads and writes the same config and links as the GUI, making it ideal for SSH sessions, headless machines, and terminal-only workflows. This skill teaches agents to drive skm correctly: init, enable/disable, doctor, and fix. Note that skm does not author, search, or translate skills — those live in the GUI or elsewhere.
Runs skm init to create ~/.skills-manager/config. and detect installed AI tools on the machine; runs skm list -- to show hub skills and each tool's link status; runs skm enable/disable <skill> --for <tool> to create or remove symlinks (Windows: symlink → junction → tracked copy); runs skm adopt to move real skill directories in tool folders into the hub and relink them (with --dry-run preview); runs skm doctor to detect tools and report sync issues, and skm fix --yes to repair them.
- Set up full skill management on a GUI-less SSH/headless server using skm init + adopt
- Diagnose 'a skill is not showing up in Claude Code/Codex' — doctor locates the issue, fix repairs broken links
- Adopt skill directories already sitting in a tool's folder into the hub so they stop bypassing central management
- Enable a skill for only the specific tool the user named instead of every detected tool
- Users who want to search or install skills.sh marketplace skills — skm has no marketplace; use npx skills or the GUI
- Users who want to author SKILL.md content, tag, translate, or manage custom tools via CLI — these are GUI-only
- Users who want to relocate the hub directory — the CLI hardcodes ~/.skills-manager/skills
How do you install this skill?
- Caution: `skm adopt --` without `--yes` still applies changes (the confirmation prompt is skipped); always pass `--yes` explicitly and preview with `--dry-run` (human mode) first.
- `skm doctor --` always exits 0 — gate scripts on `issues_count`, not the exit code; `fix --yes --` may print a result on stdout while exiting 1.
- The CLI binary must be downloaded from overseas sites (GitHub Releases / official website) with no checksum or signature verification guidance in the repo; verify integrity yourself. The app is not Apple-notarized.
- Enabling a skill writes symlinks (or tracked copies on Windows) into AI tool directories — a real filesystem side effect; use only with informed consent.
- This was a static review with no execution; all behavior claims come from the docs themselves and must be verified by the user.
- Shell / CLI
- Local filesystem
Skills Manager desktop app or skm binary (GitHub Releases)terminal / shell access
Install the Skills Manager desktop app or a standalone skm binary first; the skill is auto-installed into the hub by skm init.
Desktop app (macOS via Homebrew)
brew tap jiweiyeah/tap
brew install --cask jiweiyeah/tap/skills-managerStandalone CLI (GitHub Releases)
# Download skm-<target>.tar.gz (.zip on Windows) from https://github.com/jiweiyeah/skills-manager/releases, extract, put on PATH
skm --versionCompanion skill installs automatically with CLI setup
skm init --How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Initialize skm on this SSH server and adopt the skills already sitting in my tool directories
- My ab-testing skill isn't showing in Claude Code — diagnose and fix it with skm
- Enable the skills-manager-cli skill for Codex only, don't touch other tools
- List all skills and their per-tool link status as JSON
Triggered when a user or agent needs to manage skills from a terminal, SSH session, CI job, or headless machine. Typical flow:
skm init -- # first run, idempotent
skm adopt --dry-run # preview directories to adopt
skm adopt --yes -- # adopt and relink
skm list -- # confirm state
skm doctor -- # health check (gate on issues_count, not exit code)
skm fix --yes -- # repair reported issuesGotchas: enable/disable have no --; write commands take an advisory lock (retry once on conflict); errors in JSON mode go to stderr as {"error":"..."}; fix and adopt require --yes to actually apply.
What are this skill's strengths and limitations?
- GUI and CLI share the same config and symlinks, so they are interchangeable — good for dotfiles scripts and headless setups
- Designed for agent callers: -- output, idempotent init, dry-run previews, advisory locks against concurrent writes
- Open-source MIT, cross-platform on macOS/Windows/Linux, supports 32 AI tools plus custom tools
- Symlink sync avoids file duplication; on Windows a normal account works via three fallback strategies
- The CLI cannot author, edit, search, translate, or marketplace-install skills — the GUI is required for those
- Tool enabled/detected flags cannot be flipped from the CLI; you must use the GUI Tools page
- skm is not a standalone npm package — you must install the desktop app or grab a binary from Releases
- Tools installed after init may keep showing 'not installed' until the GUI re-detects them; the macOS build is not Apple-notarized
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Skills Manager CLI (skm) Skill this page | 59 · Recommended | ★ 1.1k | 1mo ago | MIT |
| Engram Persistent Memory Protocol | 68 · Recommended | ★ 7.7k | 8d ago | MIT |
| Vibe Replay | 56 · Use with care | ★ 38 | 7d ago | MIT |
| Multica Skill Importer | 56 · Use with care | ★ 54k | 6d ago | NOASSERTION |
| OpenSkills Universal Skills Loader | 54 · Use with care | ★ 11k | 9mo ago | NOASSERTION |
The skills.sh npx skills CLI is a different, unrelated tool: skm only manages the local hub and symlinks, not marketplace search/install; the GUI adds marketplace, translation, and tagging that the CLI does not implement.
How did FollowSkills review this skill?
The skill only drives a local CLI: no network calls, no credential access; scope is limited to ~/.skills-manager and symlink management in tool skill directories; write commands use an advisory lock; the Windows fallback chain (symlink→junction→tracked copy) records provenance transparently. Deductions: `adopt` moves real user directories, and `--` without `--yes` still skips confirmation and applies (the doc admits this), so a false trigger by an agent mutates state; the binary must be downloaded from a website or Releases with no checksum/signature guidance, so supply-chain integrity is unverifiable statically.
SKILL.md and references/.md document JSON contracts, exit codes, the dry-run/apply matrix (including the counterintuitive `adopt -- --dry-run` silent output), lock conflicts, and error messages in exceptional, self-consistent detail — near-exemplary diagnosability at the documentation level. Capped at 10: no executed key-path reproduction, and no committed tests covering the skm commands themselves (workflow tests only cover the publish script).
The description states precise triggers (managing skills, missing skill, running skm) and explicit negative boundaries (not npx skills, skill-creator, marketplace); Out-of-scope and Which-tool tables narrow scope further; cross-platform (with Windows permission notes) and headless scenarios are well covered. Deductions: the core dependency must be fetched from overseas sites (GitHub Releases, ClawHub, official website) with no China mirror or verification means; many `--` semantic traps raise correct-usage friction despite accurate triggering.
Well-layered docs (SKILL.md → .md → tools.md), clear boundary declarations, install locations, prefix-collision tables, MIT license, and version-gated publishing (publish-skill.yml). Deductions: the skill itself has no changelog/version field; its update path depends on ClawHub publishing tied to app versions; publisher is an unverified individual; README contains sponsorship/ad content.
High marginal value over manually creating symlinks across 32 tools; machine-readable output suits agent automation; docs are directly usable. However, static review cannot verify actual output correctness (junction fallback, adopt move behavior) — completion evidence is author-claimed only, so capped at 7, scored 6.
Auditable primary material exists: source repo, MIT license, real CI workflows (build.yml builds and attaches skm; publish-skill.yml with version gating and unit tests). But CI tests cover only the publish script, not the skm behavior the skill claims; the JSON contracts have no independent reproduction or third-party corroboration — thin multi-source verification, capped at 5, scored 4.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →