Build a Self-Growing Knowledge Base with WorkBuddy & Obsidian: Step-by-Step Guide

Core Difference: Traditional vs. Self-Growing Knowledge Base

Traditional note-taking tools like Obsidian or Notion often become "collect-and-forget" repositories. Stored articles remain isolated Markdown files with no connections, no versioning of ideas, and no way to reconcile conflicting viewpoints. A self-growing knowledge base, by contrast, fuses new input with existing knowledge: the AI reads the existing wiki, merges updates into relevant pages, creates new pages only for genuinely new concepts, and preserves conflicting perspectives with source attribution.

Tool Preparation

Two tools are required:

Obsidian – local Markdown storage, bidirectional links, graph view, fully offline.

WorkBuddy – acts as the knowledge-base maintainer: reads new material, extracts key points, cross-references existing wiki pages, and writes incremental updates.

Create a new empty Obsidian vault (e.g., MyKnowledge). Do not manually create any folders; let WorkBuddy handle the structure.

Minimal Three-Layer Directory Structure

Deep hierarchies confuse AI maintenance. Only three top-level folders are needed: raw/ – read-only archive of original articles, clipped web pages, meeting notes. Never modified. wiki/ – AI-maintained structured knowledge, the long-term asset. output/ – temporary reports, outlines, or analysis results.

Send the following initialization prompt to WorkBuddy to create the directories and write the governance rules into AGENTS.md:

请把当前 Obsidian 知识库配置成一套可以长期增量维护的 LLM Wiki。创建 raw、wiki、output 三个目录，并把维护规则写入 AGENTS.md。

raw 只保存原始资料，不修改原文。
wiki 由你维护，每次处理新资料前，先读取 wiki/index.md 并搜索相关页面，已有主题直接补充，没有合适页面再新建。写入 wiki 的重要结论必须标注对应的 raw 文件路径，不同来源说法不一致时同时保留。
output 用来保存临时报告和分析结果。
每次处理完成后更新 wiki/index.md 和 wiki/log.md。

请先创建目录和规则，不要处理现有文件。

Verify that raw/, wiki/, output/ exist and inspect AGENTS.md – it serves as the constitution ensuring read-only raw files, incremental updates, and mandatory source tracing.

First Ingestion

Add a single test article (converted to Markdown via browser clipper) into raw/. Then send the ingestion prompt:

请读取 raw 中刚加入的资料，并按照 AGENTS.md 增量维护 wiki。先查找已有相关页面，再决定更新还是新建。完成后更新 index.md 和 log.md，并告诉我本次新增了什么、修改了什么、哪些内容存在冲突或还不能确认。

WorkBuddy reads the source, generates structured concept pages under wiki/, refreshes wiki/index.md (global index), and records detailed changes in wiki/log.md. The original file in raw/ remains untouched.

Incremental Evolution Verification

The real test is the second article on the same topic but with new perspectives. Place it in raw/ and run the same ingestion prompt. Observe two behaviors:

Existing concepts are enriched in-place – no duplicate files created.

Core conclusions carry relative paths to both source files; conflicting views are preserved side-by-side with attribution.

If both hold, the knowledge-fusion loop is working.

Advanced Option: claude-obsidian

For large-scale, long-term maintenance (tens of thousands of words), the open-source claude-obsidian toolkit provides engineered workflows with built-in validators: init – bootstrap a new knowledge base. adopt – take over an existing vault. wiki-ingest – ingest new material. wiki-query – precise Q&A with provenance. wiki-lint – periodic audit for orphan pages and broken links.

Choose the prompt-driven approach for daily reading notes and initial experiments; graduate to claude-obsidian when the knowledge base grows complex and requires automated consistency checks.

Self-Check Checklist

Strict isolation of raw and wiki. raw/ archives originals; wiki/ holds distilled knowledge. Raw files are never modified.

Traceability of key facts. Every core assertion in wiki/ must include the relative path of its source file.

Topic merging. New material preferentially updates existing pages to avoid information silos.

Clear change log. log.md records every ingestion’s additions, modifications, and conflicts for auditability.

Synchronized index. index.md always reflects the current knowledge map.

Information that does not connect is merely digital clutter. Let each input trigger self-comparison and incremental iteration to transform fragments into a true second brain.