skills-analytics

Skills Analytics for Claude Code

Stop guessing which skills matter. This Claude Code plugin tracks every skill invocation across all your projects and scores each one. You know what to keep, improve, or remove.

Demo

https://github.com/user-attachments/assets/0eeb34f8-f564-4022-a1c3-d1062d5825f9

You have skills accumulating over time but there's no way to know which ones deliver value and which are dead weight. Install this plugin once, and it logs every skill invocation and nested file access via Claude Code hooks across every project, then provides a local dashboard with time-normalized analytics to answer: which skills should I keep, improve, or remove?

What it does

Data collection (runs automatically via Claude Code hooks):

• Logs every skill invocation with metadata (name, source, scope, timestamp)
• Tracks which nested files within skills are accessed (progressive disclosure coverage)
• Snapshots your skill inventory on each conversation start to detect additions and deletions

Analytics dashboard (local Django server, on-demand):

• Frequency — most/least used skills, sortable
• Adoption — how quickly new skills gain traction
• Usefulness scoring — composite score factoring in usage rate, decay, and engagement depth
• Trends — time-series of invocations by day/week/month
• Structure coverage — per-skill file tree showing which nested files are triggered vs. dormant

Why time-normalization matters

Skills are added at different times. A skill installed yesterday with 5 invocations (5/day) is healthier than one installed 3 months ago with 30 invocations (0.3/day). Raw counts are misleading.

The same applies to nested files — a reference doc added last week shouldn't be penalized alongside files that have existed for months.

The usefulness scoring model handles this:

Factor	What it measures
Time-normalized usage rate	Invocations per day since install, with a 7-day grace period for new skills
Usage decay detection	Is recent usage declining vs. lifetime average?
Engagement depth	What fraction of a skill's nested files actually get triggered?

Skill sources tracked

Source	Scope	Location
Folder-based	User	~/.claude/skills/
Folder-based	Project	.claude/skills/
Plugin-based	User	Installed for you across all projects
Plugin-based	Project	Installed for all collaborators on a repo
Plugin-based	Local	Installed for you, in one repo only

Deleted skills are preserved in the dashboard as "removed" with full historical data.

Installation

Requires Python 3.10+ and uv.

1. Add the marketplace and install

claude plugin marketplace add hardness1020/claude-skills-management
claude plugin install skills-analytics

Installs at user scope by default, tracking skill usage across all your projects. Hooks are auto-registered, no manual settings.json editing needed.

2. Restart Claude Code

Start a new Claude Code session for the hooks to take effect. Every prompt you submit will now snapshot your skill inventory, and every skill invocation or file read will be logged.

3. Launch the dashboard

Use the /skills-analytics-dashboard skill in any Claude Code session. It starts a local server and opens the dashboard at http://localhost:8787.

Architecture

Claude Code Session (any project)
    |
    |-- UserPromptSubmit         -->  inventory_snapshot.py --> SQLite
    |-- PreToolUse (Skill|Read)  -->  log_event.py  -->  SQLite
    |
    |           Django Dashboard (localhost:8787)
    |               |
    |               +-- reads from same SQLite

Hooks are short-lived Python scripts using only stdlib (sqlite3, json, sys). No Django import overhead in the hot path. The dashboard is a separate process that reads the same SQLite database.