ClaudeRadar v2

A Claude Code plugin that reads every session in ~/.claude/projects/ and grades how well you collaborate with AI. 9 dimensions across 3 categories — Communication, Engineering, Outcome. Returns a free-form AI diagnosis and at-least-5 pastable improvement prompts. 100% local. Dashboard view.

🌏 中文版 · 📖 Methodology · ⚖️ License

What's new in v2

v2 is a major rewrite. v1 reports are not back-compat.

• 9 dimensions across 3 categories — Communication / Engineering / Outcome. Replaces v1's flat 6-dim model. The new Engineering category evaluates how you leverage the platform: Skills, MCP, Subagents, CLAUDE.md, Plan mode.
• Project profile auto-detection — every project is auto-classified (one-shot / feature-build / long-running / learning). A 3-message bugfix is no longer unfairly compared to a 50-session feature. Dimensions can be N/A per profile.
• Free-form AI diagnosis — replaces v1's MBTI. You get a 150-word collaboration profile, a core strength + bottleneck paragraph, and a cross-dimension reading.
• Pastable prompt rewrites — every suggestion ships with a concrete prompt you can copy directly into your next session.
• CWD detection — /claude-radar first asks "analyze this project?" based on your current directory. No more scrolling through 100 projects.
• Dashboard layout — left nav + right content, with scroll-tracked highlighting. Replaces the v1 single-scroll report.

What the report includes

Run /claude-radar and get a single-file HTML dashboard:

Overall grade — S through D, paired with your project profile so you know what scale you're being judged on.

Three category scorecards — each scoring 0-100 with its 3 dimensions:

  COMMUNICATION  78  [A]    ENGINEERING  62  [B]    OUTCOME  80  [A]
  Lock-On        82          Toolcraft    68         Efficiency   88
  Scene Setting  71          Architecture N/A        Proof Check  62
  Steering       81          Tempo        56         Completion   90

Diagnosis — core insight (strength + bottleneck), a free-form collaboration profile, and a cross-dimension pattern reading. This is the part most users find valuable.

Improvement plan — 5-7 suggestions (always at least 5, even for high-scoring users), each with cited evidence, a pastable prompt rewrite, and the expected score impact. High-scoring users get "level-up moves" instead of corrective ones.

Drill-downs — detailed reasoning for each of the 9 dimensions, plus a Toolcraft & Architecture detail panel showing skills used, MCP servers, Plan mode usage, CLAUDE.md status, and more.

Install

Step 1 — Add the marketplace:

/plugin marketplace add LeifDiao/claude-radar

Step 2 — Install the plugin:

/plugin install claude-radar@claude-radar-marketplace

Alternative (local):

git clone https://github.com/LeifDiao/claude-radar.git ~/claude-radar
claude --plugin-dir ~/claude-radar

Use

/claude-radar

1. ClaudeRadar detects your current working directory and asks "analyze this project?"
2. Say yes, or pick from the recent-10 list
3. Wait ~30 seconds for parsing + diagnosis
4. The dashboard opens in your browser

Requirements

• Claude Code with plugin support
• Node.js 18+ (already ships with Claude Code)
• No npm install, no build step, no server

Privacy

Your session data stays on your machine:

• Everything runs locally — no network calls
• No API key, no telemetry, no cloud
• ~/.claude/projects/ is read-only
• Reports write to ~/.claude-radar/reports/
• CLAUDE.md / memory / agents detection only reads filesystem metadata, never file contents (except CLAUDE.md size)

How scoring works

Two-layer model:

1. Scoring — deterministic baseline formula + bounded Claude ±15 adjustment with cited evidence. Total variance ~±3 points.
2. Diagnosis — independent qualitative pass. Free-form 150-word collaboration profile, core diagnosis, cross-dimension reading. Replaces MBTI.

Fairness mechanisms:

• Project profile drives category weighting and N/A dimensions
• Density-based confidence — a 5-message project with high signal density no longer gets unfairly shrunk
• Efficiency as a first-class signal — "3 messages, 5 files edited" is recognized as high efficiency, not penalized as low volume

👉 Read the full Methodology

Transparent rubric

All scoring rules live in data/rubric.json:

• 9 dimension definitions, baseline formulas, applicability rules
• 3 category groupings with per-profile weight tables
• Grade thresholds (S/A/B/C/D)
• Diagnosis and suggestion specifications
• Density-based confidence scaling

If you want scoring to match your team's standards, edit this file. Claude re-reads it every run.

Project structure

claude-radar/
├── .claude-plugin/plugin.json        # plugin manifest (v2.x)
├── skills/analyze/
│   ├── SKILL.md                      # cwd detection + 9-dim scoring + diagnosis
│   └── scripts/
│       ├── list-projects.mjs         # scan projects + cwd match
│       ├── parse-project.mjs         # facts extraction (tool/skill/MCP/CLAUDE.md detection)
│       └── render-report.mjs         # JSON → HTML
├── viewer/template.html              # dashboard report template
├── data/rubric.json                  # 9-dim formulas + profile weights + diagnosis spec
└── docs/
    ├── METHODOLOGY.md                # methodology (English)
    └── METHODOLOGY_zh.md             # methodology (Chinese)

About 250 KB. Zero runtime dependencies.

License

CC-BY-NC-4.0 — free for non-commercial use.

Built for people who care about the quality of AI collaboration.