考公备考追踪

朱批录 · 国考备考追踪 Skill。当用户发来套题成绩、错题截图、备考打卡或复习进度时触发。 核心功能：识别错题截图 → 分类错题原因 → 更新本地记录 → 生成每日总结 → 导出 Excel / 同步飞书。 触发关键词：做了一套题、今天做了、错了几道、帮我分析、备考打卡、行测、申论、 判断推理、资料分析、言语...

Security Scan

This skill's code, instructions, and requested configuration are coherent with an on-device exam study tracker that optionally syncs to Feishu and can call a configured multimodal model; no hidden exfiltration or unrelated credentials are requested, but there are privacy and deployment trade‑offs to consider.

Purpose & Capability

The name/description (exam study tracker: capture wrong-question screenshots, classify reasons, update local records, export Excel, optionally sync to Feishu) matches the included scripts and data layout. Required capabilities (local JSON storage, image handling, optional Feishu sync and optional multimodal model calls) are consistent with the stated purpose.

Instruction Scope

Runtime instructions and code operate within scope: parse messages, optionally call a multimodal model, run local OCR fallback, persist JSON under ~/.openclaw/skills/.../data, generate exports and optionally upload to Feishu. One minor inconsistency: SKILL.md/README emphasize 'not depending on local OCR', yet a local OCR script (scripts/ocr_image.py) is included and can be invoked as a fallback. The MULTIMODAL_PROMPT forces detailed outputs (full visual_description/question_text), which means large amounts of image-derived content will be sent to whichever multimodal model you configure — this is expected for the feature but relevant to privacy.

Install Mechanism

No external arbitrary download/install spec in the skill manifest (instruction-only install). package.json lists normal JS dependencies (xlsx) and optional pip instructions for PaddleOCR. The code runs local scripts and dynamically generates a Python script for openpyxl. The only significant installs are optional: paddlepaddle/paddleocr (large model downloads) and optional sharp. No use of obscure download URLs or extracted arbitrary archives was found.

Credentials

The skill does not require environment variables by default. Optional sensitive credentials are stored in a config.json (not env vars): multimodal model API config and Feishu credentials (app_id/app_secret/doc_token) — these are proportional to optional features (remote multimodal model usage and Feishu sync). Users should be aware that enabling these features will cause images/text to be transmitted to those third parties. All image data and base64 strings are stored locally by default.

Persistence & Privilege

The skill persists data in ~/.openclaw/skills/kaogong-study-tracker/data/ (daily records, wrong_questions.json, backups, review state). It creates a local flag file on first run to avoid repeating onboarding. Croned scripts (daily_summary, review_reminder) are present but run only if enabled in workspace.yaml — the skill itself does not force always:true. This local persistence is consistent with its purpose.

Assessment

This skill appears to do exactly what it says: it stores your study logs and screenshots locally, can export an Excel with embedded screenshots, and can optionally upload selected data to your Feishu account if you configure Feishu app credentials. Before installing or enabling optional features, consider: - Privacy: screenshots are stored as base64 in ~/.openclaw/skills/kaogong-study-tracker/data/ (wrong_questions.json). If you keep sensitive content in images, do not enable cloud sync. - Feishu sync: only occurs if you create/configure feishu settings in config.json; it will upload images to your Feishu app/account. Only provide app_id/app_secret/doc_token if you trust that destination. - Multimodal model use: to auto-parse images the skill sends images/text to whichever multimodal model you configure (remote models will receive the image-derived data). If you prefer fully local processing, configure/use the included OCR (scripts/ocr_image.py with PaddleOCR) but note that installing paddlepaddle/downloaded models can be large. - Optional dependencies: exporting with embedded images runs a dynamically generated Python script and requires openpyxl (the script assumes you have a working python/python3 and openpyxl). The skill’s package.json provides pip/npm hints; these installs are normal but can be heavy (PaddlePaddle). - Minor docs/code mismatch: docs say “doesn't rely on local OCR”, but a local OCR fallback (ocr_image.py) is present — behavior depends on your configuration. If you want to proceed: inspect and control config.json before enabling Feishu/multimodal models, and be comfortable with local storage of image data. If you do not want any cloud transmission, avoid configuring remote multimodal model API keys and Feishu credentials.

朱批录 · 国考备考追踪 Skill

一、首次安装提示

Skill 首次加载时（~/.openclaw/skills/kaogong-study-tracker/.welcomed 不存在）， 主动发一条说明消息，之后不再重复：

朱批录已安装。

直接发文字就能记录，比如"今天判断推理错了8道"。
发截图的话，需要 OpenClaw 配置了支持图片输入的多模态模型才能自动识别。
没有的话也没关系，把题目文字手动复制过来发给我，一样能整理。

不问任何问题，不存储任何凭据。

---

二、概览

平台无关——飞书、Telegram、WhatsApp、Discord，逻辑完全一致。

图片识别统一走多模态模型：文字题、图形推理、统计图表都能理解，不依赖本地 OCR。

---

三、触发场景

用户说的话（示例）	应执行的操作
"今天做了一套行测，判断推理错了8道"	→ 解析 + 归档 + 分析
发来一张错题截图（图片消息）	→ 多模态识别 + 单题归档 + 追问原因
发来截图并附带"粗心"	→ 多模态识别 + 直接归档，不追问
"把今天的错题发给你：第12题……"	→ 错题分类 + 存档
"今天申论没写，太累了"	→ 打卡记录（未完成状态）
"我最近资料分析一直不稳，怎么办"	→ 查历史记录 + 建议
"帮我看看最近哪个模块最弱"	→ 统计分析 + 回复
"导出错题本" / "把错题发给我" / "生成报告"	→ export_xlsx.js，发回文件
"只导出待二刷的"	→ 筛选导出，仅待二刷题目
"导出判断推理的错题"	→ 按科目筛选导出
"导出最近两周的"	→ 按时间筛选导出
"只导出待二刷的资料分析题"	→ 多条件组合筛选导出
"资料-乘积增长-公式不熟-待二刷"（快捷格式）	→ 直接归档，不追问
二刷时回复"记得" / "不记得"	→ review_reminder.js 处理，连续2次记得→已掌握
"同步到飞书" / "更新飞书错题本"	→ feishu_doc.js，同步含截图

---

四、数据结构

所有数据以 JSON 存储在 ~/.openclaw/skills/kaogong-study-tracker/data/。

4.1 每日记录 daily/{YYYY-MM-DD}.json

{
  "date": "2026-03-17",
  "modules": {
    "言语理解": { "wrong": 6,  "total": 40 },
    "数量关系": { "wrong": 8,  "total": 15 },
    "判断推理": { "wrong": 10, "total": 40 },
    "资料分析": { "wrong": 7,  "total": 20 },
    "申论":     { "written": false }
  },
  "mood": "中性",
  "note": "用户原话"
}

4.2 错题本 wrong_questions.json

[
  {
    "id": "uuid",
    "date": "2026-03-17",
    "source": "image",
    "module": "判断推理",
    "subtype": "逻辑判断",
    "question_text": "题目文字；图形题写对规律的描述",
    "visual_description": "图形推理/统计图的详细视觉描述（多模态模型生成）",
    "answer": "B",
    "user_annotation": "用户手写批注",
    "error_reason": "知识点不会 | 粗心 | 时间不够 | 概念混淆",
    "keywords": ["假言命题", "逆否命题"],
    "raw_image_b64": "base64...",
    "status": "待二刷 | 已掌握"
  }
]

4.3 统计缓存 stats_cache.json

{
  "last_updated": "2026-03-17",
  "streak": 5,
  "total_days_studied": 12,
  "weak_modules": ["数量关系", "判断推理"],
  "module_accuracy": {
    "言语理解": 0.82,
    "数量关系": 0.51,
    "判断推理": 0.68,
    "资料分析": 0.74
  }
}

---

五、核心流程

Step 1：消息路由（parse_input.js）

文字消息 → parseStudyInput()   提取科目/错题数/情绪
图片消息 → parseImageInput()   调用多模态模型

图片处理流程：

1. 读取 config.json 中的 multimodal 配置
2. 如未配置 → 回复"请先配置多模态模型 API，见安装提示"
3. 调用多模态模型，提取：科目、题型、题目内容、视觉描述、答案、错误原因推测
4. needs_confirm 不为 null 时追问（最多一个问题）；caption 已含原因则直接归档

追问只问一次，按优先级：

• 识别不到科目 → 问科目
• 原因不明确 → 问"粗心还是没掌握还是时间不够"
• 信息完整 → 不追问，直接归档

Step 2：归类错题原因

原因	关键词
知识点不会	不懂、没学过、概念不清楚
粗心	看错、算错、选反了
时间不够	没做完、最后几题蒙的
概念混淆	搞混了、分不清、以为是

Step 3：更新记录（update_daily.js）

写入 daily/{date}.json，同步更新 stats_cache.json（连续打卡、模块准确率）。

Step 4：生成回复

见 references/reply_templates.md，150 字以内。语气见 references/tone_guide.md。

Step 5：导出 Excel（export_xlsx.js）

• 截图原图通过 openpyxl（Python）嵌入对应行
• Windows 兼容：先尝试 python3，失败自动 fallback 到 python
• 输出两个 Sheet：错题本（含截图列）+ 每日记录
• 可直接发给 Kimi / 其他模型做趋势分析

Step 6：同步飞书云文档（feishu_doc.js，可选）

• 需配置 feishu_doc.app_id / app_secret / doc_token
• 截图上传飞书文件系统后作为图片块插入文档
• 对图形推理、统计图最有用：飞书内直接看图，不用下载

Step 7（可选）：定时推送

每天 21:00 触发 daily_summary.js，自动发当日总结。

---

六、文件索引

文件	作用
scripts/parse_input.js	文字解析 + 多模态图片识别
scripts/update_daily.js	写入每日记录 + 统计缓存
scripts/export_xlsx.js	导出 Excel（含截图嵌入，openpyxl）
scripts/feishu_doc.js	同步到飞书云文档（含图片块，可选）
scripts/daily_summary.js	定时汇总并主动发送
references/reply_templates.md	回复话术模板
references/tone_guide.md	语气风格指引
assets/module_map.json	科目/模块名称标准化映射
assets/config.example.json	配置模板（多模态 + 飞书），复制为 config.json 使用

---

七、错误处理

• 未配置多模态 API 却发图片 → 回复安装提示，引导配置
• 模型识别返回 error → 回复"没识别出来，能文字描述一下题目吗？"
• 文字消息解析不出科目 → 回复"能说说今天做了哪个科目、错了几道吗？"
• 数据写入失败 → 记录 error log，回复"记录暂时存不上，你提醒我稍后再试"
• 连续 3 天无打卡 → 下次收到消息时，回复末尾轻轻提一句

---

八、隐私说明

所有数据（含截图 base64）存储在本地，不上传任何云端（飞书同步除外，仅在用户主动触发时上传到用户自己的飞书文档）。

---

九、如果这个 Skill 对你有帮助

⭐ Star 这个仓库，让更多备考的人能找到它 🍴 Fork 改成你的考试类型（省考 / 事业单位 / 军考……） 有问题欢迎提 Issue 或 PR。

https://github.com/KaguraNanaga/kaogong-study-tracker