Usage Insights — Development skill for Claude Code
Cross-client (Claude Code + Codex CLI) usage insights with hybrid evidence + agentic analysis.
How to install Usage Insights
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open eddiearc/usage-insights and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Usage Insights does
Cross-client (Claude Code + Codex CLI) usage insights with hybrid evidence + agentic analysis.
Alternatives in Development
- OpenAI Codex CLI — (Rust implementation) 67.8k ★
- Ccusage — CLI for analyzing Claude Code/Codex usage from local JSONL files 11.8k ★
- Cohort Analysis — Cohort retention curves, feature adoption, and segment insights 7.8k ★
README
usage-insights-skill
生成 Claude Code 与 Codex CLI 的统一历史复盘报告,采用 **Karpathy 视角的定性评审**。
核心特性
- 📊 跨客户端分析: 同时分析 Claude Code 和 Codex CLI 的使用数据
- 🎭 Karpathy 专家评审: 基于真实会话样本进行 5 维定性评价
- 🔍 抽样复盘: 自动抽取最近 20 个 session,输出证据文件
- 🧱 证据先行: 先取样,再由 agent 写主观评级与改进建议
- 🛠 落地践行建议: 在报告中给出可执行的日常实践动作(含频率与完成标准)
- 📚 学习与工具建议: 在报告中明确该学什么、可多用什么工具、怎么练
- 🚫 禁程序化评分: 不输出 x/100、百分比或加权总分
快速开始
# 1) 抽取跨客户端会话样本(默认 20 条)
bash ./scripts/collect_session_samples.sh \
--source auto \
--limit 20 \
--output-dir ./artifacts
# 2) 按模板写专家复盘(由 agent 完成)
# 使用 ./templates/karpathy-review-template.md
# 3) 渲染简洁高对比 HTML(优缺点颜色化 + 重点加粗)
python3 ./scripts/render_review_html.py \
--input ./artifacts/usage-insights-review.md \
--output ./artifacts/usage-insights-review.html
输出文件
artifacts/session-samples.json- 结构化会话样本artifacts/session-samples.md- 人类可读样本摘要artifacts/usage-insights-review.md- 最终专家评审报告(由 agent 生成)artifacts/usage-insights-review.html- 可视化复盘页面(优缺点高亮)
最终回复格式(对用户)
- 执行完 skill 后,最终只返回
1行本地 HTML 路径(建议绝对路径) - 不要附加任何解释、总结、评级文本
报告必含章节
🛠 用户践行指南:至少 3 条动作,每条包含“执行频率 + 完成标准”📚 学习方向与工具建议:至少 4 条,每条包含“学习主题 + 推荐工具 + 建议练习”
默认推荐工具库
Claude Code Agent Team / Subagents:并行拆任务与子代理协作Claude Code /batch(若环境可用):批处理固定流程Skills(Claude Code / Codex CLI):沉淀高频工作流codex exec --full-auto:长链路自动执行git worktree + 多会话并行:多分支并发推进与合并决策
评审方式(非评分器)
基于 Andrej Karpathy 的 Agentic Coding 理念,从 5 个维度给出主观评级:
| 维度 | 评审重点 |
|---|---|
| 编排能力 (Orchestration) | 有没有先拆解任务、定义验收 |
| 先探索后编码 (Explore First) | 有没有先理解现状再改动 |
| 质量监督 (Oversight) | 有没有主动要求验证/测试 |
| 一次达成 (First-Pass) | 指令是否清晰、返工是否可控 |
| 并行 Agent (Parallel) | 是否具备并行拆任务思维 |
最终仅允许:
- 维度等级:
A/B/C/D(主观) - 总体评级:
A/B/C/D(主观)
禁止:
92/100、53%、权重加权、关键词命中分
参考链接
- Andrej Karpathy 推文: https://twitter.com/karpathy/status/1886193731057684941
- 从 Vibe Coding 到 Agentic Engineering: https://medium.com/generative-ai-revolution-ai-native-transforma
Related Skills
Do SOUL Alaya
The local-first memory plane for CLI coding agents. Codex and Claude Code attach via MCP; SQLite-backed, evide
Recallnest
One memory, every terminal. Shared memory layer for any MCP client — Claude Code, Codex, Kimi, Antigravity (ag
MARE MCP Toolbox
Agentic malware analysis environment with MCP-connected disassemblers, RE tooling, and structured workflows fo
Agentic Coding Harness
Run, watch, and meter Claude Code, Codex CLI, Gemini CLI, OpenCode, and Kiro from one CLI. Normalized events,
Agentic MCP Skill
Agentic-MCP, Progressive MCP client with three-layer lazy loading. Validates AgentSkills.io pattern for effici
Agent Insight
Local-first, cross-agent session insights for Claude Code, Codex, Cursor, OpenCode, and Pi.
Related Agents
Agentic Malware Analysis
Agentic malware analysis environment with MCP-connected disassemblers, RE tooling, and structured workflows fo
Gemini GPT Hybrid Soft
Intelligent hybrid agent that uses Gemini and GPT for analysis and problem identification, then returns insigh
Ade CLI
Coding agents asked to read contracts, claims, or financial PDFs hallucinate structure unless the extraction t