lms-plugin-skills

A 1:1 clone of Claude's internal skill system, built as an LM Studio plugin.

How It Works

Claude reads a list of available skills at the start of every context, then uses a view tool to read the relevant SKILL.md file before working on tasks that skill covers. This plugin replicates that system exactly for LM Studio.

The Skill System - Three Components

1. Prompt Preprocessor Before every message, the plugin scans the skills directory and injects an <available_skills> block into the prompt. The model sees this and knows which skills exist, their descriptions, and their file paths - exactly like Claude's system prompt injection.

2. Tools

Tool	Purpose
list_skills	List all available skills with names and descriptions
read_skill_file	Read any file within a skill directory (defaults to SKILL.md)
list_skill_files	Explore the full file tree of a skill directory
read_file	Read the contents of any file in the workspace outside of the skills directory
write_file	Create or overwrite a file with exact contents, avoiding shell escaping issues
patch_file	Find and replace a specific string inside an existing file for surgical code edits
delete_file	Delete files or directories: requires explicit recursive=true to delete non-empty directories (prevents accidental folder deletion), with guards to handle non-existent paths gracefully.
move_file	Move files/directories: matches shell mv behavior (moves source into existing destination directories), and automatically creates missing parent directories for new nested paths.
rename_file	Rename files/directories within their current directory only: validates no path separators in the new name, and fails instead of overwriting existing files to prevent accidental data loss.
append_to_file	Atomically append content to files (OS-level atomicity safe for logs), and creates missing parent directories (can also create new files with initial content).
create_directory	Idempotent mkdir -p wrapper that creates nested directories without errors if they already exist.
list_directory	List directory contents: walks one level deep by default (recursive=false, lean output), or caps at 10 levels when recursive=true to avoid overwhelming output. Reuses shared tree formatting to match list_skill_files output.
get_current_directory	Returns home directory, current working directory, and platform info to eliminate path guessing across operating systems.
run_command	Execute shell commands, run scripts, and interact directly with the local system

3. Explicit Skill Activation

You can activate a skill directly from your message using /skill-name notation:

/bug-fix there is a null pointer error in this function: …
use /git-commit to push my changes
/docx write a report summarising the findings

Any token matching /[a-z][a-z0-9._-]* (regex) is treated as an explicit activation. The preprocessor resolves it before the request reaches the model:

• The matching SKILL.md body is expanded inline into a <skill_context> block - the model does not need to call read_skill_file for it.
• The named skill is treated as the highest-priority context; all other message text is secondary task payload.
• Quoted strings, code snippets, globs, and command-looking text in the message are passed through to the skill untouched.
• run_command must not be used for exploration when a skill is explicitly activated.
• If the name does not match any installed skill, the model is instructed to call list_skills with that name as a query before proceeding.

Multiple activations in one message are supported (/skill-a and /skill-b are both expanded). Explicit activation takes priority over auto-inject; the <available_skills> block is not appended on the same turn.

4. Persistent Settings LM Studio does not save plugin settings across new chats. This plugin solves that by writing settings to ~/.lmstudio/plugin-data/lms-skills/settings.json - the skills path and all configuration survive chat resets when only skillsPath field set as default.

---

Skill Directory Structure

A skill is any subdirectory inside your skills folder that contains a SKILL.md file.

~/.lmstudio/skills/          <- default skills directory
├── docx/
│   ├── SKILL.md             <- entry point (required)
│   ├── scripts/
│   │   └── helper.py
│   └── templates/
│       └── base.docx
├── pptx/
│   ├── SKILL.md
│   └── editing.md
└── my-custom-skill/
    ├── SKILL.md
    └── skill.json           <- optional: override name/description

skill.json (optional)

Place a skill.json in any skill directory to override its display name and description:

{
  "name": "My Custom Skill",
  "description": "Use this skill when the user asks to do X, Y, or Z.",
  "tags": [
    "data analysis",
    "csv",
    "statistics",
    "charts",
    "visualisation",
    "pandas",
    "trends",
    "dataset"
  ]
}

If absent, the plugin uses the directory name and extracts the description from the first paragraph of SKILL.md.

---

Settings

Setting	Default	Description
Auto-Inject Skills List	On	Injects skills block into every prompt
Max Skills in Context	15	Max skills listed in each injected block
Skills Directory Path	(empty)	Custom path to skills directory

Skills Directory Path

• Empty - uses the last saved path (or ~/.lmstudio/skills on first run)
• default - resets the saved path back to ~/.lmstudio/skills
• Any absolute path - saves that path and uses it immediately

Settings (including the skills path) are written to disk and survive new chat sessions.

---

Local Development

cd lms-plugin-skills
bun install
lms dev

---

Default Skills Path by Platform

The default path ~/.lmstudio/skills resolves to:

Platform	Path
Windows	C:\Users\<you>\.lmstudio\skills
macOS	/Users/<you>/.lmstudio/skills
Linux	/home/<you>/.lmstudio/skills

---

Model Workflow

Auto-inject (default)

1. User sends a message
2. Preprocessor fires - scans skills dir, injects <available_skills> block
3. Model reads the block and recognises a relevant skill
4. Model calls read_skill_file("skill-name") → receives full SKILL.md content
5. SKILL.md may reference other files → model calls list_skill_files then read_skill_file with specific path
6. Model follows the skill's instructions to produce high-quality output
7. Model uses read_file, patch_file, and write_file to execute the actual coding and file management tasks

Explicit activation (/skill-name)

1. User includes /skill-name anywhere in their message (prompt)
2. Preprocessor detects the tokens, resolves the skill, and expands its SKILL.md into a <skill_context> block - no tool call needed
3. Model receives the expanded skill body as highest-priority context alongside the user's task payload
4. If the skill name is not found, the model calls list_skills to locate it before proceeding
5. Model follows the skill's instructions to produce high-quality output

License

• Apache 2.0