wiki-knowledge-agent
Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. When a user pastes a link or text (in ANY channel — Discord, Slack, CLI, etc.), fetch+verify it, classify it (useful vs ad), translate+summarize it into the target language, save it to the wiki, and optionally alert on time-sensitive information. Also acts as a RAG channel: agents search the wiki for prior knowledge. Use whenever the user pastes content for archiving, asks to save a link/bookmark, asks '알림 있어?', 'any alerts?', or asks the agent to look up knowledge saved in the wiki.
atukunare · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add atukunare/wiki-knowledge-agent --skill wiki-knowledge-agent技能说明
name: wiki-knowledge-agent description: "Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. When a user pastes a link or text (in ANY channel — Discord, Slack, CLI, etc.), fetch+verify it, classify it (useful vs ad), translate+summarize it into the target language, save it to the wiki, and optionally alert on time-sensitive information. Also acts as a RAG channel: agents search the wiki for prior knowledge. Use whenever the user pastes content for archiving, asks to save a link/bookmark, asks '알림 있어?', 'any alerts?', or asks the agent to look up knowledge saved in the wiki." category: note-taking triggers:
English
- "save this"
- "save this link"
- "bookmark"
- "bookmark this"
- "add to wiki"
- "remember this"
- "archive this"
- "store this"
- "clip this"
- "note this"
- "paste this"
- "wiki"
- "knowledge base"
- "any alerts?"
- "check alerts"
- "notifications"
- "save for later"
한국어
- "북마크"
- "링크 저장"
- "링크 저장해줘"
- "저장해줘"
- "위키에 저장"
- "위키에 넣어줘"
- "붙여넣기 정리"
- "인박스"
- "알림 있어?"
- "알림 확인"
Wiki Knowledge Agent
Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. Works on any agent platform (Hermes, Claude Code, Codex, Cursor) with zero external dependencies beyond curl + Python stdlib.
When to Use
- User pastes a link or text into any chat channel and expects it to be captured.
- User says "저장해줘", "링크 저장", "북마크", "save this", "add to wiki".
- User asks "알림 있어?" — read the local alert queue and report.
- An agent needs prior knowledge from the wiki (RAG usage — search
wiki_root). - A scheduled job runs the ingest/notify cycle.
Design Principles
- Platform-independent — works with a single model (no multi-agent setup required) or many.
- Zero external dependencies —
curl+ Python stdlib only. No webhook/app required for core function. - Local-first alerts — alerts always land in a local file + CLI output. Optional webhook/email/ntfy if configured.
- Any-channel input — a message arriving in any channel is eligible. The onboarding default channel is "the current chat".
- LLM does the reasoning — the scripts fetch/verify/structure; the model classifies ads, translates, and summarizes.
Configuration
Config file: ~/.config/wiki-knowledge-agent/config.yaml (default), overridable via WIKI_AGENT_CONFIG env var.
Run onboarding to create it interactively:
python3 scripts/onboarding.py
Example config (full reference: templates/config.example.yaml):
wiki_root: "~/wiki"
target_language: "ko"
translate: true
input:
sources: ["any"] # "any" = every channel/CLI; or a specific list e.g. ["discord", "slack"]
default_channel: "current" # channel used when no preference is set (onboarding default)
notify:
tier0: true # always: local file + CLI
webhook: "" # optional: Slack/Discord/Telegram webhook URL
email: "" # optional: SMTP DSN
ntfy_topic: "" # optional: ntfy.sh topic
alert_on:
security: true
account_expiry: true
project_risk: true
interest: false
Onboarding flow (interactive questions)
- Wiki root path → default
~/wiki - Input channels → choose from known platforms (discord/slack/weixin/telegram/cli…) or "any"; if the user doesn't pick, default is "current chat" (
default_channel: current). Content pasted into other channels is still judged and saved (never drop input because it arrived in the "wrong" channel). - Target language → default user's native language (e.g.
ko). - Alert channel → if the user has a webhook/email, capture it; otherwise default is the current chat (the agent reports alerts in the same conversation). Alerts can be added later anytime: just ask the model "알림 채널 추가해줘" — the LLM updates the config.
- Write
config.yamland print the summary.
Ingest Workflow (when a message arrives)
[input] pasted text or link in any channel
│
▼
1. VERIFY (script: scripts/ingest.py --verify <url-or-file>)
- Link: curl -s -I -L → HTTP status; 2xx/3xx = valid, 404/410/timeout = dead
- Text: mark as unverified text
│
▼
2. CLASSIFY (LLM, guided by references/ad-detection.md)
- 📢 Ad: self-promotion is the core intent, "try it" pitch, link-dump without substance
- 📚 Useful: real methodology/data/tool with transferable value
- Output: label + one-line reason
│
▼
3. PRE-SAVE FEEDBACK (LLM — before saving, add value from what you already know)
a) Project/context match: if the content relates to anything you know about
the user (their projects, goals, interests, past notes), say so in one line:
"📌 이거 <프로젝트>에 유용하겠네요" / "📌 This looks useful for <project>"
b) Date/event cue: if the content has a date-related element (holiday, birthday,
anniversary, deadline, event) and you know the relevant date, offer ONE question:
"📅 <이벤트>(<날짜>) <N>일 전에 알려줄까요?" / "📅 <event> is in <N> days — remind you?"
- If the user says yes → record it (e.g. add to config important_dates, or note it)
- If unsure of the exact date → don't guess; skip the offer
c) Keep it light: at most ONE feedback line + ONE date offer. Never block saving on it.
│
▼
4. TRANSLATE + SUMMARIZE (LLM)
- Foreign content → target_language
- ≤5-line summary + 3–5 key points
- If already in target_language, summarize only
│
▼
5. SAVE (script: scripts/ingest.py --save ...)
- <wiki_root>/knowledge/inbox/YYYY-MM-DD-<topic>.md
- Frontmatter: date, source_url, source_channel, classification (ad/useful), language
- Tag: related project or "일반"
│
▼
6. ALERT? (LLM decides against alert_on rules + script: scripts/ingest.py --notify)
- If alert-worthy → append to <wiki_root>/alerts/notifications.md
+ CLI output "[알림] …" + optional webhook/email
- Else → silent save; reply "✅ 인박스: <fi
阅读完整来源 ↗