SKILL.md — Obsidian Dual-Layer Knowledge Base System
> Version: 0.2 | Updated: 2026-05-22
> Skill Name: obsidian-knowledge-base
> Tags: knowledge-management, obsidian, rag, agent-memory, daily-pipeline
> Install: npx clawhub@latest install obsidian-knowledge-base
1. Skill Overview
Turn Obsidian into a Dual-Layer Knowledge Base for AI Agents, supporting the full pipeline: daily incremental writes → structured indexing → context injection.
- INBOX Layer: Raw material entry (articles/conversations/logs), unstructured
- WIKI Layer: Structured knowledge (concept cards + Q&A cards), directly callable by AI
Core Value: Solve the "knowledge base paralysis" problem — knowledge gets written in, but the AI starts from scratch every time because it doesn't know what's in the base.
2. Directory Structure
Obsidian Vault/
├── INBOX/ # Input layer (raw materials)
│ ├── articles/ # Articles (daily scheduled writes)
│ ├── conversations/ # Conversation records
│ └── 日志/ # Work logs (date-indexed)
└── wiki/
├── qa/ # Q&A knowledge cards
│ ├── index.md # Q&A index (by category)
│ └── Q-分类-摘要.md
└── concepts/ # Concept cards (deep structure)
├── index.md # Concept card index
├── 概念-主题名.md
└── ...
3. Knowledge Base Scale & Context Strategy
| Layer | File Count | Raw Size | How It Enters Context |
|---|---|---|---|
| INBOX/articles | Daily incremental | Raw material | Not injected; searched on demand |
| wiki/qa | ~200-500 cards | ~500KB-1MB | Daily 6KB summary injected |
| wiki/concepts | ~20-100 cards | ~50-100KB | Searched on demand |
Context Strategy (Progressive Disclosure):
- At startup: Run wiki_daily_brief.py → generate 6KB category summary → inject into context
- During tasks: Specific question → tdai_memory_search / wiki_search → read raw snippet (~700 chars)
- Forbidden: Do not pour full wiki content into context (session limit: 200KB)
4. Naming Conventions
4.1 INBOX/articles/
【Category】YYYY-MM-DD-Title.md
Example: 【Industry Research】2026-05-22-Some Industry Deep Dive Summary.md
Category Tags: 城市更新 / 物业管理 / 土地市场 / 楼市分析 / 城投转型 / 政策速递 / 行业研究 / 他山之石
4.2 wiki/qa/
Q-Category-ContentSummary.md
Example: Q-IndustryResearch-WhatAreKeyPointsOfCommercialPropertyLeaseManagement.md
4.3 wiki/concepts/
Concept-TopicName.md
Example: Concept-CityRenewal.md
5. Card Templates
5.1 Q&A Card Template (frontmatter)
---
type: qa
date: 2026-05-22
title: Q-Category-ContentSummary
tags: [Category]
scenarios: [Report Writing/Article Writing/Policy Research]
keywords: [keyword1, keyword2]
operation: knowledge-query
---
## Question
(Specific question)
## Answer
(Complete answer, cite the source text, do not generalize)
5.2 Concept Card Template (frontmatter)
---
type: concept
date: 2026-05-22
title: Topic Name
tags: [Category, Subcategory]
source: Source
summary: One-sentence core conclusion
---
# Concept: Topic Name
## When to Use
When encountering the following keywords/scenarios, prioritize this card:
- Keyword 1
- Keyword 2
## Core Conclusion (One Sentence)
## Detailed Content
## Source/Evidence
6. Filter Rules (Mandatory)
Content PROHIBITED from writing to wiki/qa/ and wiki/concepts/:
| Category | Examples | Correct Destination |
|---|---|---|
| Work logs / daily matters | "Today's completed work", "Core completion items" | INBOX/日志/ |
| OpenClaw operations config | "How to batch process documents", "cron config" | skills/SKILL.md |
| Memory/session mechanisms | "OpenClaw memory scheme", "session structure" | memory/ (non-wiki) |
| Scheduled tasks/Cron | "How to write cron tasks" | skills/SKILL.md |
| Technical troubleshooting | "API Key invalid troubleshooting" | skills/SKILL.md |
Content ALLOWED in wiki:
Any structured knowledge refined through analysis and processing, including but not limited to:
- Industry/domain knowledge (concepts, models, cases, data)
- Policy interpretation and applicable scenarios
- Core conclusions and methods from research reports
- Standards and operational guidelines
- Cross-domain comparative analysis
Rule of Thumb: If a user can directly use it to write a report or make a decision, it's knowledge.
7. Daily Pipeline
05:00 ── INPUT END ──────────────────────────────
│
│ cron: Daily news scraping
│ → Scrape news → write to INBOX/articles/
│ → Push Word summary to relevant people
│
│ Note: Do NOT write to wiki/ directly; 8:30 handles it
│
08:30 ── OUTPUT END (Knowledge Base Injection) ──
│
│ cron: Daily knowledge base injection
│ → Scan INBOX/articles/ for today's new files
│ → Write to wiki/qa/ (extract Q&A pairs)
│ → Write to wiki/concepts/ (extract concept cards)
│ → Filter out operational content (see Section 6)
│ → Run wiki_daily_brief.py to update daily summary
│ → Push summary to owner (X new cards today, Z total in base)
│
│ At Session Startup (automatic):
│ → Run wiki_daily_brief.py
│ → Inject 6KB category summary into context
│ → AI knows what's available in the knowledge base today
8. Retrieval Flow (Per Task)
Receiving an analysis/writing task
→ wiki_daily_brief already injected at startup (6KB summary)
→ wiki_search(topic keywords) → match Q&A cards
→ High match → use the answer directly
→ Low match → wiki_search(subtopic) → read concept card
→ Still insufficient → search INBOX/articles/ raw materials
→ Last resort → web_search
9. wiki_daily_brief.py Script
#!/usr/bin/env python3
"""
wiki_daily_brief.py
Run at first session startup each day. Generates a wiki knowledge base
daily summary so the model knows what's available in the knowledge base.
Output: ~/.openclaw/workspace/memory/wiki_daily_brief.md
"""
WIKI_QA_DIR = "/path/to/Obsidian/wiki/qa"
OUT = "~/.openclaw/workspace/memory/wiki_daily_brief.md"
# Script logic:
# 1. Scan all .md files in wiki/qa/ (excluding index.md)
# 2. Extract frontmatter tags, count by category
# 3. Take top 3 per category as examples, generate ~6KB summary
# 4. Write to OUT, ready for session startup injection
10. Daily Summary Format
# Wiki Knowledge Base Daily Summary
> Generated: 2026-05-22 | Total: 230 Q&A Cards
---
## Available Knowledge Cards (By Category)
### Industry Research (83 cards)
- **What is the core content of「Some Report」?** (Source: `Q-IndustryResearch-SomeReport.md`)
- ...
- ...and 80 more
### City Renewal (66 cards)
- ...
## Usage Guide
To query specific card content, call `tdai_memory_search` or `wiki_search`.
11. Installation
# Method 1: ClawHub
npx clawhub@latest install obsidian-knowledge-base
# Method 2: Manual
cp -r obsidian-knowledge-base ~/.openclaw/skills/
12. Use Cases
- Knowledge workers who need AI to write reports, analysis, or official documents
- Teams that want AI to have "memory" instead of starting from scratch every time
- Scenarios with daily incremental material inputs (news, reports, data)
- Teams that want knowledge base and AI context to work together, not in silos
