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. With a query, scores and ranks skills by relevance using BM25-inspired search across name, tags, description, and SKILL.md body content.
read_skill_file	Read any file within a skill directory (defaults to SKILL.md). Accepts either a skill name or an absolute path.
list_skill_files	Explore the full file tree of a skill directory. Accepts either a skill name or an absolute path.

3. Explicit Skill Activation

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

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.
• If the name does not match any installed skill, the model is instructed to call 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.

skill.json (optional)

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

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

---

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)

• User sends a message
• Preprocessor fires - scans skills dir, injects <available_skills> block
• Model reads the block and recognises a relevant skill
• Model calls read_skill_file("skill-name") → receives full content

Explicit activation (/skill-name)

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

Default Skills — Data Analysis Scripts

The data-analysis skill ships with three Python scripts in scripts/ that use only the Python standard library (no pandas dependency):

Script	Purpose
profile.py	Full dataset profile: types, nulls, stats, top values
correlations.py	Pearson + Spearman correlation matrix with tie handling
time_series.py	Date-aware trend analysis, resampling, rolling averages

For larger datasets (>100k rows), pandas is recommended for better performance.

License

• Apache 2.0

License