离谱了兄弟们，天天折腾 AI 做“第二大脑”，结果笔记库全被 AI 幻觉填成了垃圾回收站？这玩意真有人做出来了，直接从底层把你拯救了。

GitHub 刚上新的开源项目 ai-content-kb。这不是什么割韭菜的套壳应用，而是一套开源（MIT 协议）的 AI 辅助知识库参考架构。专门给硬核写手和研究党量身定制，刚刚冒头热乎着呢。

核心就干几件事，直戳痛点：

审查优先：AI 只能当底层黑工，所有生成内容必须人工过审，绝对不乱改原稿。

三区隔离：原始素材、AI 生成废料、最终成品严格分开，强迫症狂喜。
在线工具

工具百搭：纯 Markdown 底座，能完美串联 Obsidian、Codex 和各类 Agent。

来源可溯：内容带有结构化关系，生成的废话一秒查出源头，彻底杜绝数据污染。
在线工具

这波有点狠，思路比一堆吹上天的 SaaS 软件清晰多了，白嫖党直接拿去抄作业就行。

GitHub 指路：https://github.com/mrbear1024/ai-content-kb…

mrbear1024/ai-content-kb

Source: https://github.com/mrbear1024/ai-content-kb

This project was inspired by Andrej Karpathy’s LLM Wiki pattern: immutable raw sources, a persistent and compounding Markdown wiki maintained by an LLM, and a repository schema such as AGENTS.md that defines how the agent should ingest, query, and maintain knowledge.

ai-content-kb extends that idea with explicit provenance roles, a first-class published-product layer, review-first staging, typed machine-readable relationships, content-creation workflows, and safe legacy-vault migration.

Most knowledge bases mix personal thinking, external evidence, published work, and AI-generated text in the same pile. This template gives each a clear role—and gives AI agents rules for working without silently polluting originals or publication workflows.

Originals stay trustworthy. The wiki stays readable. The graph stays queryable. AI output stays reviewable.

The screenshot shows a mature vault using this architecture. A fresh clone starts with a small synthetic example and grows as you add reviewed notes, links, and tags.

Why this project

Quickstart

git clone https://github.com/mrbear1024/ai-content-kb.git
cd ai-content-kb

No application is required to read the vault. Use plain Markdown, open the root as an Obsidian vault, or use the intended Codex workflow below.

Open in Codex

1. Open the Codex desktop app.
2. Select + in Projects.
3. Choose Use an existing folder.
4. Select the cloned repository root.
5. Start a new task and say:

Inspect the project structure, read the knowledge-base rules,
make no changes, and list the available workflows.

Codex reads repository AGENTS.md guidance before starting work. The root instructions then route it to the rest of this project’s rules. See the official Codex documentation.

The phrases below are natural-language workflows defined by this repository, not native slash commands. No plugin is required.

Talk to your knowledge base

How it works

flowchart LR
    R["raw/\nOwner input"]
    S["sources/\nExternal evidence"]
    P["products/\nReviewed output"]
    T[".kb/staging/\nAI candidates"]
    W["wiki/\nHuman interface"]
    G[".kb/links/\nMachine graph"]

    R -->|ingest| T
    S -->|ingest| T
    P -->|backfill| T
    T -->|human review| W
    T -->|human review| G
    W -->|compose| P
    G -->|retrieve evidence| P

Read the architecture rationale and graph schema for details.

Core workflows

Capture and index

add to knowledge base: this attachment is an external source.
Preserve provenance, check existing aliases, and keep generated wiki and graph data in staging.

Review .kb/staging/wiki/ and .kb/staging/links/ before promotion.

Create with AI

Compose from the knowledge base:
Write an article for AI product managers about Context Engineering.
Separate my judgments from external claims, cite repository paths,
and create a source plan and outline before drafting.

Unreviewed prose goes to .kb/staging/drafts/, accepted work moves to raw/drafts/ for development, and only publication-ready work belongs in products/.

Read the content creation guide.

Migrate an existing vault

Migrate legacy knowledge base: inventory the provided directory in read-only mode.
Do not move, copy, delete, or modify files.
Propose a 20–50 file pilot and stop for review.

Migration follows backup → inventory → mapping review → pilot → validation → reviewed batches → cutover.

Read the migration guide.

Explore in Obsidian

Open the repository root with Open folder as vault, then select Graph view in the left ribbon. The included .obsidian/graph.json colors nodes by raw, sources, products, and wiki.

Obsidian visualizes Markdown links and tags. Typed .kb/links/*.yaml relationships are a separate machine layer and do not automatically become Obsidian lines.

Review-first lifecycle

The deciding factors are review status, intended use, human ownership, and publication readiness—not whether AI helped write the text.

Documentation

Project scope

This repository is a reference architecture and working template, not a hosted knowledge-management application. Automation is intentionally minimal: validate the content roles and review workflow with files first, then add manifests, search, embeddings, or a graph database only when real queries justify them.

Obsidian and Codex are optional interfaces. The durable system is the repository itself.

Contributing and support

• Found a problem or have a proposal? Open an issue.
• Pull requests improving the information model, examples, schemas, lint rules, and documentation are welcome.
• Do not submit private notes, copyrighted captures, credentials, absolute home paths, or generated databases containing personal metadata.
• Before publishing a fork, use the public release checklist.

Maintained by mrbear1024 and contributors.

License

MIT. This project is not affiliated with or endorsed by Obsidian or OpenAI.