Claude Skill Manager (csm)

Manage Claude Code skills across projects from a single place.

The Problem

Claude Code requires skills to live flat inside each project's .claude/skills/ directory. If you have a library of skills organized in subdirectories, or skills shared across multiple projects, you have to copy them manually into every project — and keep them in sync.

# What you have (organized by topic):
~/skills-library/
  data-analysis/
    backtest/SKILL.md
    fetch-data/SKILL.md
  productivity/
    meeting-prep/SKILL.md

# What Claude Code needs (flat, per project):
~/project-a/.claude/skills/backtest/SKILL.md
~/project-a/.claude/skills/fetch-data/SKILL.md
~/project-b/.claude/skills/meeting-prep/SKILL.md

csm solves this by creating symlinks from your organized skill library into each project's .claude/skills/ directory. You maintain skills in one place, and deploy them anywhere.

Agent-Friendly by Design

csm is built to be used by both humans and LLM agents:

• csm schema outputs a complete JSON document describing all commands, concepts, config format, and keybindings — designed to be injected into an agent's context
• --json flag on every query command for structured, parseable output
• Errors are always JSON on stderr (even without --json), following the same structured format as GWS
• No interactive prompts in JSON mode — agents can script csm safely

An agent can bootstrap itself with:

csm schema | head -c 4000  # inject into system prompt
csm --json sources          # discover what's available
csm --json targets          # discover where to deploy
csm install my-skill --to my-project  # deploy

Core Concept

csm connects sources (where skills live) to targets (where Claude Code looks for them):

	Description	Example
Source	A directory containing skills (SKILL.md) or a Claude Code marketplace	~/skills-library, ~/code/my-org/*
Target	A project with a .claude/ directory	~, ~/code/my-project
Install	A skill deployed from a source into a target	symlink or claude plugin install

Demo

Installation

With Homebrew (macOS)

brew install aclemen1/tap/claude-skill-manager

With uv (any platform)

uv tool install claude-skill-manager

With pip

pip install claude-skill-manager

Quick Start

# 1. Create a config file
csm init

# 2. Edit config to add your source and target paths
cat ~/.config/claude-skill-manager/csm.toml

Example csm.toml :

source_paths = ["~/code/my-org/*", "~/skills-library"]
target_paths = ["~", "~/code/my-org/*"]

Paths support glob patterns. For source_paths, the wildcard marks where skill directories are expected:

• ~/skills-library — exact path; skills are direct subdirs (~/skills-library/my-skill/SKILL.md)
• ~/code/my-org/* — each matched dir may be a skill; source root = parent (~/code/my-org)
• ~/code/my-org/*/* — two levels deep; source root = one level up from the match

For target_paths, each resolved directory is checked for a .claude/ subdirectory.

# 3. Explore what's available
csm sources          # list all discovered skills
csm targets          # list all projects with .claude/
csm list             # unified inventory with install state

# 4. Install a skill
csm install my-skill --to my-project

# 5. Or use the interactive TUI
csm

CLI Commands

Command	Description
csm	Launch the interactive TUI (default when no subcommand)
csm sources	List all discovered sources (local skills, marketplace plugins)
csm targets	List all targets with install counts
csm installs	List all current installs (symlinks, plugins, orphans)
csm install SKILL --to TARGET	Install a local skill into a target via symlink
csm uninstall TARGET	Remove skill symlinks from a target
csm adopt SKILL --from TARGET --to SOURCE	Adopt an orphan: move it into a source library and replace it with a symlink
csm list	Unified inventory with install state per item
csm diagnostics	Detect per-target name collisions and issues
csm updates	Detect stale plugin cache
csm init	Create a default config file
csm schema	Output JSON schema for LLM/agent consumption

JSON output

All query commands support --json for machine-readable output:

csm --json sources    # JSON array of sources with items
csm --json targets    # JSON array of targets with counts
csm --json installs   # JSON array of all installs
csm --json list       # JSON inventory with install state
csm --json diagnostics --all  # JSON diagnostics

Errors are always returned as structured JSON on stderr:

{
  "error": {
    "code": 1,
    "message": "No items matched 'nonexistent'.",
    "reason": "notFound"
  }
}

LLM/Agent integration

csm schema  # outputs full JSON documentation for LLM consumption

The schema command outputs a structured JSON document with all commands, concepts, config format, diagnostic types, and TUI keybindings — designed to be injected into an LLM context for tool use.

TUI Keybindings

Navigation

Key	Action
j / k	Move down / up
Enter	Expand / collapse (fold)
l	Expand
h	Collapse, or go to parent
L	Expand all under cursor
H	Collapse all under cursor
Tab / Shift+Tab	Cycle panels

Selection and Install

Key	Action
Space	Select (switch to toggle mode on other panel)
x	Toggle install / uninstall
a	Apply pending changes
d	Delete selected pending change
Esc	Cancel all pending changes

Preview

Key	Action
p	Preview SKILL.md (source panel: any skill; target panel: orphans)
e	Edit in $EDITOR (inside preview modal, local skills only)

Modals and Actions

Key	Action
A	Adopt orphan (move it into a source library, replace with symlink)
s	Settings editor
D	Diagnostics (conflicts + stale cache)
?	Help
Ctrl+P	Command palette (change theme, etc.)
r	Refresh (rescan all sources)
q	Quit

Theme selection previews each theme in real time as you navigate the list. The chosen theme is persisted in csm.toml.

Diagnostics

csm detects issues per target — two skills with the same name only conflict if they're both active in the same project:

Type	Severity	Description
user-user	ERROR	Two local skills with the same name in the same target
user-plugin	WARNING	A local skill shadows a Claude Code plugin
orphan-plugin	WARNING	An unmanaged skill copy coexists with a plugin
cross-marketplace	WARNING	Same skill name from different marketplaces
scope-overlap	INFO	A plugin is installed in both user scope and project scope
mp-cache	INFO	Normal: marketplace catalog and installed cache

Configuration

Config file: ~/.config/claude-skill-manager/csm.toml

# Enable Claude Code marketplace plugin discovery (default: true)
plugins = true

# Glob patterns for skill source directories.
# Semantics depend on whether the pattern contains a glob wildcard:
#   - Exact path "~/skills-library": the directory itself is the source root;
#     skills are its direct subdirectories (~/skills-library/my-skill/SKILL.md).
#   - Glob "~/code/my-org/*": each matched directory may be a skill;
#     csm checks for SKILL.md directly inside it, and the source root is its
#     parent (~/code/my-org/my-skill/SKILL.md → source = ~/code/my-org).
#   - Glob "~/code/my-org/*/*": same, two levels deep
#     (~/code/my-org/lib/my-skill/SKILL.md → source = ~/code/my-org/lib).
source_paths = [
    "~/skills-library",
    "~/code/my-org/*",
]

# Glob patterns for target directories
# Each resolved directory is checked for .claude/ presence
target_paths = [
    "~",
    "~/code/my-org/*",
]

# Theme preference (set via Ctrl+P in TUI)
# theme = "monokai"

How it works

The flat deployment problem

Claude Code scans .claude/skills/ for skill directories, each containing a SKILL.md. It expects a flat structure — no nesting:

my-project/.claude/skills/
  backtest/SKILL.md       ← Claude Code sees this
  fetch-data/SKILL.md     ← and this

But as a developer, you want to organize your skills by topic, team, or library:

~/skills-library/
  data-analysis/backtest/SKILL.md
  data-analysis/fetch-data/SKILL.md
  productivity/meeting-prep/SKILL.md

Symlink-based deployment (managed by csm)

csm bridges the gap with symlinks. When you install a skill into a target, csm creates a symlink in the target's .claude/skills/ directory pointing back to the source:

~/project-a/.claude/skills/
  backtest → ~/skills-library/data-analysis/backtest
  fetch-data → ~/skills-library/data-analysis/fetch-data

The skill lives in one place (the source library), but is visible to Claude Code in every project where it's installed. Edit the skill once, every project picks it up.

Marketplace plugins (managed by Claude Code)

For Claude Code marketplace plugins, csm provides a unified view of all installed plugins across all your projects and scopes (user/project). Install and uninstall operations are delegated to the claude plugin CLI. The toggle operates at the plugin level (not individual skills), since Claude Code installs entire plugins.

Plugins that contain no skills (e.g. hook-only or agent-only plugins) are also shown — csm reads plugin.json metadata to display their name and description.

In the TUI, plugins show their install scope with visual indicators:

Indicator	Meaning
● green	Installed directly in the selected target (toggleable)
● blue + "via user scope"	Inherited from user scope (toggleable to add project scope)
● green + [user+project]	Installed in both scopes (toggleable for project scope)
○ dim	Not installed (toggleable to install)
[user] / [project] / [user+project]	Scope badge in normal mode (not target-specific)

Orphans

Skills found in .claude/skills/ that are neither symlinks to known sources nor installed via plugins are flagged as orphans (shown with ? in the TUI). These are typically skills that were copied or created manually.

You can adopt an orphan to bring it under csm management: adoption moves the skill directory into a configured source library and replaces it with a symlink, so it becomes a proper managed skill. Use csm adopt from the CLI or press A on an orphan node in the TUI.

Development

# Clone and install
git clone https://github.com/aclemen1/claude-skill-manager.git
cd claude-skill-manager
uv sync

# Run tests
uv run pytest

# Run the TUI locally
uv run csm

# Run with coverage
uv run pytest --cov=skill_manager

License

MIT