Case 01428

Skill knownasnaffy/prompthound/dataset/case_01428

A fast, offline static risk analysis CLI for AI agent skill files. Detects malicious instructions, steganographic payloads, and dangerous capability chains.

Install

npx -y skills add knownasnaffy/prompthound --skill case_01428

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Collect and organize a personal knowledge base from URLs (web/X/WeChat) and screenshots. Use when the user says they want to save an URL, ingest a link, archive content to KB, tag/classify notes, store screenshots, or search their saved knowledge in Telegram. Supports WeChat via a connected macOS node when cloud fetch is blocked.

SKILL.md

3.7 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it

Summary

Ingest: web URLs, X/Twitter links, WeChat Official Account links (mp.weixin.qq.com), and screenshots
Store: writes to a shared KB folder with per-item content.md + meta.json and a global index.jsonl
Organize: tag-first classification with richer tags (e.g. #agent, #coding-agent, #claude-code, #mcp, #rag, #prompt-injection, #security, #pricing, #database)
WeChat: cloud fetch may be blocked; when a macOS node (e.g. Reed-Mac) is online, prefer node-side fetch to improve success rate; otherwise create a placeholder entry
Search: designed to support Telegram Q&A / search flows on top of the index and content

把用户发来的链接/截图沉淀到共享知识库（KB），并做标签化整理。

默认 KB 位置

KB Root（可改）：/home/ubuntu/.openclaw/kb
索引：kb/20_Inbox/urls/index.jsonl
每条内容目录：kb/20_Inbox/urls/<YYYY-MM>/<item>/content.md + meta.json

目标：先入库不丢，再迭代“摘要/标签/检索”。

你要做的事（按输入类型）

1) 普通网页 / X(Twitter) / 公众号 URL 入库

运行脚本：

python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/ingest_url.py "<URL>" --tags "#optional" --note "context"

行为：

自动识别来源（web/x/wechat）
优先用 r.jina.ai 抽取正文（无需登录）
公众号遇到风控会写占位条目：status=blocked_verification + tag #needs-manual
对同一 URL 做 key 去重（已存在则跳过）

WeChat 更高成功率（推荐路径）

当云端抓取命中“环境异常/验证”时：

如果有已连接的 macOS 节点（例如 Reed-Mac）且该节点能访问该文章，可用 nodes.run 在节点上执行抓取（requests+bs4），然后写入 KB。
注意：这条路径依赖节点在线与网络环境；无法承诺 100%。

2) 截图/图片入库（含 OCR 文本）

脚本：

python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/ingest_image.py /path/to/image.jpg \
  --text-file /path/to/ocr.txt \
  --title "..." --tags "#ai #product" --note "..."

说明：

ingest_image.py 负责“落盘+索引”。OCR 可用：
- 本机 tesseract（若安装了 tesseract-ocr + chi_sim）
- 或用多模态 LLM 抽取文字后写入 --text-file

Telegram 里直接问（检索）

推荐先用脚本（本机/服务器）：

python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/search_kb.py --q "claude code" --limit 10
python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/search_kb.py --tags "#claude-code #coding-agent" --limit 20
python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/search_kb.py --source wechat --since 7d --q "Elys"

公众号待补抓队列（占位条目）

python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/wechat_backlog.py --limit 30

周报/主题报告候选清单（给 LLM 写总结用）

python3 /home/ubuntu/.openclaw/skills/knowledge-base-collector/scripts/weekly_digest.py --days 7 --limit 30

重要注意事项（安全/隐私）

截图/网页可能包含 token/验证码/密钥：入库前应做脱敏（替换为 REDACTED）。
公众号抓取受风控影响：建议允许“占位入库”，后续再补全。