Claude Radar
A Claude Code plugin that reads your project conversation records and grades how well you collaborate with AI. 9 dimensions across 3 categories — Communication, Engineering, Outcome. Returns a free-form AI diagnosis, at-least-5 pastable improvement prompts, and a professional, readable HTML dashboard. 100% local.
🌏 中文版 · 📖 Methodology · 🖥 Live preview · ⚖️ License
中文简介： 一款 Claude Code 插件，读取你的项目对话记录，从「沟通力 / 工程力 / 成效」三个层面、9 个维度评估你和 AI 的协作质量。输出一份 AI 撰写的协作诊断、至少 5 条可直接粘贴的改写 prompt，以及一个专业可读的 HTML dashboard。所有计算 100% 本地完成。完整中文文档 → README_zh.md
Key features
🎯 Reads your actual sessions, not synthetic prompts. Most "AI productivity" tools test you on contrived examples. Claude Radar analyzes the real conversations you've had with Claude — every directing, correcting, and confirming message, plus every tool call you made.
💬 AI writes you a coaching note, not just a scoreboard. Beyond the 9 scores, you get a 150-word collaboration profile that describes how you specifically work with AI, a one-paragraph core diagnosis pairing your strongest strength with your most critical bottleneck, and a cross-dimension reading that explains how your behaviors combine. Every claim cites evidence from your real session.
📋 Every suggestion is a pastable prompt. No "be more thoughtful" advice. Each of the 5–7 improvement suggestions comes with a concrete prompt you can copy-paste straight into your next session, plus the expected score impact and the trade-off.
⚖️ Project-aware fairness. A 3-message bugfix isn't compared to a 50-session feature build. Claude Radar auto-classifies each project (one-shot / feature-build / long-running / learning) and applies different category weights and N/A rules. Density-based confidence prevents short-but-substantive sessions from being unfairly shrunk.
🛠 Scores how you use the platform, not just how you talk. The Engineering category specifically measures Skills, MCP servers, Subagents, CLAUDE.md, Plan mode, and custom commands — the leverage most users underuse. Not using advanced tools is fine; using them poorly (retry loops, plan-then-abandon) is what hurts.
🔒 100% local, zero telemetry. Read-only access to your local Claude Code project records. No API key, no cloud, no network calls. Bilingual (English + 中文) reports built in.
中文要点：
- 🎯 评估你真实的项目对话记录，不靠人工题库
- 💬 AI 给你写诊断信：协作画像 + 强项/瓶颈段 + 维度交叉解读，每条结论都引用真实证据
- 📋 每条建议都附可粘贴 prompt，不讲空话；附预期分数影响和取舍说明
- ⚖️ 按项目类型公平评分：bugfix 不和大项目用同一把尺子，密度高的短会话不会被打压
- 🛠 专门评估平台杠杆：Skill / MCP / Subagent / CLAUDE.md / Plan 模式 —— 不用不扣分，用得差才扣分
- 🔒 100% 本地，零数据上传，原生中英双语，输出专业可读的 HTML dashboard
完整中文版 → README_zh.md
What the report includes
Run /claude-radar and get a single-file, professional, readable 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
- Claude Radar detects your current working directory and asks "analyze this project?"
- Say yes, or pick from the recent-10 list
- Wait for parsing + diagnosis to finish (time varies with project size)
- 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:
LeifDiao
sickn33
obra
OthmanAdi
xhyqaq
Chachamaru127