contextstellar

ContextStellar — Claude Code Plugin

Real-time prompt scoring and context engineering optimization for Claude Code.

Every tool use is automatically scored across 5 dimensions. Works instantly with zero configuration — no API key, no signup, no network required.

Quick Start

# Install from marketplace
/plugin install contextstellar

# Or test locally
claude --plugin-dir ./claude-code-plugin

That's it. The plugin starts scoring immediately using the bundled local scorer.

Optional: Cloud Mode

For RLM adaptive weight learning and team dashboards, connect to the ContextStellar API:

export CONTEXTSTELLAR_API_KEY="cs_live_your_key_here"
export CONTEXTSTELLAR_PROJECT_ID="your_project_id_here"

Get credentials at contextstellar.com/connect.

How It Works

Two Modes

Mode	Config needed	Scoring	RLM Learning	Stats/Dashboard
Local (default)	None	Bundled scorer.js	No	No
Cloud	API key	Server-side + learned weights	Yes	Yes

Automatic Scoring (Hooks)

The plugin hooks into every tool use:

• PreToolUse — Scores the prompt before execution. Injects a scoring summary (grade, dimension breakdown, tips) into Claude's context so it can self-improve.
• PostToolUse (cloud only) — Reports success/failure outcomes back to ContextStellar for RLM weight learning.
• Stop (cloud only) — Captures session-end metrics (total tokens, cost).

Scored tools: Bash, Write, Edit, Agent, NotebookEdit.

Skills (Commands)

Command	Description	Requires API Key
/contextstellar:setup	Configure API key and project ID	No
/contextstellar:score <prompt>	Manually score a prompt	No (local) / Optional (cloud)
/contextstellar:stats	View scoring stats and RLM progress	Yes
/contextstellar:optimize	Get optimization recommendations	Yes

Scoring Dimensions

Dimension	Weight	What It Measures
Token Utilization	25%	Filler, politeness, intensifiers, redundancy
Structural Clarity	23%	XML tags, lists, headers, code blocks
Specificity	20%	Numbers, format keywords, examples, constraints
Content Density	17%	Lexical diversity, n-gram repetition
Cache-Friendliness	15%	Volatile data positioning, timestamps, UUIDs

RLM Flywheel (Cloud Mode)

Weights are defaults initially. As you use Claude Code, the plugin reports outcomes (success/failure) alongside scores. After 100+ events, ContextStellar learns which dimensions matter most for your project and adjusts weights automatically.

Check progress with /contextstellar:stats.

Privacy

• Local mode: All scoring happens on your machine. Zero network calls. Nothing leaves your system.
• Cloud mode: Raw prompt text is never stored — only SHA-256 hashes. stderr truncated to 200 chars.
• No API key or user content is ever logged.

Requirements

• Claude Code 1.0.33+
• Node.js (for local scorer — already present if you're using Claude Code)
• jq and curl (for cloud mode only)

Plugin Structure

claude-code-plugin/
├── .claude-plugin/
│   └── plugin.json          # Plugin manifest
├── skills/
│   ├── score/SKILL.md       # /contextstellar:score
│   ├── stats/SKILL.md       # /contextstellar:stats
│   ├── setup/SKILL.md       # /contextstellar:setup
│   └── optimize/SKILL.md    # /contextstellar:optimize
├── hooks/
│   └── hooks.json           # PreToolUse, PostToolUse, Stop
├── scripts/
│   ├── pre-tool-use.sh      # Scoring hook (local or cloud)
│   ├── post-tool-use.sh     # Outcome reporting (cloud only)
│   ├── session-stop.sh      # Session end (cloud only)
│   └── scorer.js            # Bundled local scorer (zero deps)
└── README.md

License

MIT