这次实验下来,我对个人 LLM Wiki 的最小方案有了一个新判断:不一定要先搭 Obsidian。
之前我的思路一直是基于 Obsidian 的:raw / wiki / AGENTS.md / index.md / log.md。这套逻辑没问题,尤其适合本地知识库。但如果目标只是让一个普通用户先跑通完整链路,WorkBuddy + IMA 比 Obsidian 更轻。
IMA 只做一件事:稳定获取高质量公众号内容。WorkBuddy 做剩下的:把内容放进项目资产,做知识编译、查询、回写和治理。IMA 是数据入口,真正的 LLM Wiki 在 WorkBuddy 项目里长出来。
1. 为什么 IMA 只做入口
IMA 本身更像一个 RAG 知识库——文章放进去,查询时召回相关片段,再生成回答。方便,但离我想要的 LLM Wiki 还有距离。
LLM Wiki 的关键在"持续编译"。新文章进来之后,系统要更新来源页、概念页、主题页、对比页,还要维护索引和日志。下次提问时,模型优先站在已经整理过的 wiki 层回答,而不是每次从原文片段里临时拼答案。
所以 IMA 在这里解决的是一个很现实的问题:公众号好文章怎么稳定收集。公众号内容在移动端阅读频率很高,手动导入 Obsidian 路径太长。先收藏进 IMA,再由 WorkBuddy 同步到项目资产,是目前更顺的路径。
2. WorkBuddy 项目才是主体
这套方案跑在 WorkBuddy 的"项目"概念上。
一个项目下面有动态、计划、任务、资产。当前这套 LLM Wiki 主要用任务和资产。动态和计划可以留给后续团队协作,第一版不用碰。
在 WorkBuddy 新建一个项目,比如 llm-wiki,这就是你的云端知识库工作区。
项目里的"指令"相当于以前 Obsidian 方案里的 AGENTS.md,决定 WorkBuddy 在回答、编译、查询、回写时遵守什么规则。
3. 项目指令这样写
新建项目时把下面这段放进"指令",按自己需求调整:
markdown# LLM Wiki 工作规则
## 角色
你是当前项目资产的知识编译器。请把 raw/ 中的原始资料,增量维护为 wiki/ 中可复用、可链接、可追溯的结构化知识。
## 目录
- raw/: 原始资料,只读为主
- wiki/sources/: 来源页
- wiki/concepts/: 概念页
- wiki/topics/: 主题页
- wiki/comparisons/: 对比页
- projects/: 当前项目
- workflows/: 工作流程
- index.md: 全局索引
- log.md: 操作日志
## ingest
1. 当前项目资产先读 index.md。
2. 再读 raw/ 新资料。
3. 创建或更新 sources / concepts / topics / comparisons。
4. 优先更新旧页面,避免重复建页。
5. 补充 wikilinks。
6. 更新 index.md 和 log.md。
## 查询
回答问题时,先读当前项目资产里的 index.md,再读相关 wiki 页面。wiki 不够时再回看 raw。形成新判断时,询问是否写回。
技能选项里一定要选 IMA,后面要同步 IMA 里已经收藏的文章。专家和连接器可以先不选,除非你的流程确实需要。
4. 让 WorkBuddy 自动创建目录
项目指令写好后,不需要手动建目录。在项目对话框里输入:
text按照项目指令,在当前项目资产下,生成需要的目录结构。
WorkBuddy 会读取项目指令,在项目资产根目录下创建 raw/、wiki/、projects/、workflows/、index.md、log.md 等结构。
检查一下资产目录,尤其确认 raw/ 和 wiki/ 有没有创建成功。后面所有同步和编译都依赖这两层。
5. 用自动化同步 IMA 数据
在 WorkBuddy 里创建一个自动化任务,调用 IMA 技能。这里需要你的 IMA 密钥,按引导操作即可。
自动化任务这样写:
text把 IMA 里"AI行业前沿"知识库,今天导入的文章,同步到 WorkBuddy "llm-wiki"项目资产的 /raw 目录下,保留标题、来源、平台、采集时间。不要写入 wiki,同步后更新 log.md。
提示词里要明确引用 IMA 技能,写清楚 IMA 知识库名称和 WorkBuddy 项目名称。
第一次建议手动测试。确认 WorkBuddy 能从 IMA 指定知识库里读取文章,并在项目资产 /raw 下创建文件。测试通过后,再开每日自动同步。
第一次调用 IMA 技能时需要传入 IMA 密钥:
同步结果:
6. 编译 raw 里的新增资料
raw 里有资料后,就可以做第一次编译了。
别写成"帮我整理一下知识库",太虚。直接这样:
text编译 raw 中今天新增的资料。执行前先列出计划:读取哪些 raw 文件,可能创建哪些 wiki 页面,可能更新哪些旧页面,是否更新 index.md 和 log.md。等我确认后再执行。
计划阶段输出:
确认计划后再让它执行:
编译完成后,到项目资产的 wiki/ 下看结果。重点看三件事:有没有 sources 来源页,有没有 concepts / topics / comparisons,index.md 和 log.md 有没有更新。如果只是生成了几篇摘要,还不够——真正的 LLM Wiki 要开始形成结构,页面之间最好有 wikilinks。
7. 对知识库提问
后面可以直接在这个项目里提问。例如:
text基于当前项目资产里的 wiki 内容,解释一下 LLM Wiki 和传统 RAG 的区别。请先读 index.md,再读相关 wiki 页面;wiki 不够时再回看 raw。回答后列出依据了哪些页面。
我验证过,WorkBuddy 的回答路径会按项目指令走:先看 index.md,再找相关 wiki 页面,不够再回 raw。项目指令不是摆设,确实能约束回答路径。
如果回答里出现了新的判断,可以继续让它写回:
text请把刚才的回答整理成 wiki/comparisons/LLM-Wiki-vs-RAG.md,并更新相关 topic。写回前先列出将修改或新增的资产。
有回写,知识库才会越用越厚。
8. 用定时任务治理知识库
知识库只进不管,很快会乱。再创建一个定时任务做健康检查:
text检查项目资产下,wiki/ 是否有孤立页面、重复页面、断裂链接,index.md 是否过旧,sources 是否缺少来源信息,log.md 是否完整。
生成治理报告到 projects/knowledge-health/,只提出建议,不自动删除或合并文件。
治理任务只给建议,不要自动删除或合并。知识库是长期资产,清理动作最好经过确认。
9. 用 projects 接真实工作
知识库不能只停在 raw 和 wiki。比如你要写一篇文章,直接在项目资产 /projects 下建一个项目页:
在项目资产 /projects 下创建一个项目文件,标题为:WorkBuddy LLM Wiki。
内容包括:
## 目标
写一篇公众号干货教程。
## 必须说清楚
- 目录结构
- index.md
- 资料采集
- WorkBuddy ingest
- 查询和回写
- 定时治理
## 相关页面
- LLM Wiki
- WorkBuddy
之后写文章、做选题、整理资料,都可以围绕这个 project 文件展开。这一步的价值是把知识库接进真实任务,否则它只是一个会回答问题的资料库。
10. 适合哪些场景
自媒体最对路。选题、热点、竞品文章、读者反馈、历史爆文,先进 IMA,再同步到 WorkBuddy 的 raw,定期编译到 wiki。每篇文章建一个 project,发布后把标题、阅读量、留言反馈写回。时间一长,它会越来越了解你的栏目、选题边界和标题偏好。
日常办公也适用。每个客户或项目建一个 project。会议纪要、政策通知、公众号文章、项目资料,都可以进 raw。决策写入 wiki,行动项写入 log。每天早上让 WorkBuddy 读取项目页和昨天 log,生成 morning brief;周五再生成周报。这类场景里,AI 的核心价值是工作状态不断线。
11. 几个注意点
这套方案本质上是云端 LLM Wiki。
WorkBuddy 的项目资产保存在云端,处理时会先下载到本地,处理完再上传。文件多的时候实际速度会慢一点,但好处也很直接:不用先折腾本地 Obsidian,直接用 WorkBuddy 项目能力就能跑起来。
项目资产可以下载。下载后可以直接导入 Obsidian,里面已经带有双链。这意味着 WorkBuddy + IMA 可以先作为云端方案,后面再迁移到 Obsidian 做本地归档和图谱展示。如果你想做 WorkBuddy + Obsidian 的本地笔记方案,也完全可行——底层思想一样,只是把 WorkBuddy 项目资产换成 Obsidian vault,把项目指令换成 AGENTS.md。
我现在把路线分成两种:
textWorkBuddy + IMA:云端最小方案,门槛低,适合先跑通。
WorkBuddy + IMA + Obsidian:增强方案,适合本地归档、双链图谱和长期迁移。
最后
这次实验让我不再把 Obsidian 当成必须入口。
真正的最小 LLM Wiki,核心是这条链路能不能跑通:
IMA 解决公众号内容的稳定输入。WorkBuddy 解决项目、资产、编译、查询、回写和治理。这已经足够搭一个最小可用的个人 LLM Wiki。
Obsidian 仍然有价值,但更适合做增强层。先用 WorkBuddy + IMA 跑通,再决定要不要接 Obsidian,反而更稳。
我是「超级猛」,长期折腾AI编程、Agent和自动化工作流。写Claude Code、Codex、Hermes这些工具的实战笔记,也聊AI怎么改变开发和内容生产的真实逻辑。
After this experiment, I have a new assessment of the minimal setup for a personal LLM Wiki: you don't necessarily need to set up Obsidian first.
Previously, my thinking was always based on Obsidian: raw / wiki / AGENTS.md / index.md / log.md. That logic is fine, especially for local knowledge bases. But if the goal is just to get an ordinary user through the full pipeline first, WorkBuddy + IMA is lighter than Obsidian.
IMA does only one thing: reliably get high-quality WeChat Official Account content. WorkBuddy does the rest: puts content into project assets, performs knowledge compilation, query, writeback, and governance. IMA is the data entry point; the real LLM Wiki grows inside WorkBuddy projects.
1. Why IMA is only the entry point
IMA itself is more like an RAG knowledge base: put articles in, retrieve relevant snippets on query, then generate answers. Convenient, but still far from the LLM Wiki I want.
The key to LLM Wiki is "continuous compilation." After new articles come in, the system needs to update source pages, concept pages, topic pages, comparison pages, and maintain indexes and logs. Next time you ask, the model should answer primarily from the organized wiki layer, not piece together an answer on the fly from raw article snippets every time.
So what IMA solves here is a very practical problem: how to reliably collect good WeChat Official Account articles. WeChat Official Account content is read frequently on mobile, and manually importing it into Obsidian is too long a path. Saving it into IMA first, then having WorkBuddy sync it to project assets, is currently the smoother path.
2. The WorkBuddy project is the main entity
This setup runs on WorkBuddy's "project" concept.
Under a project there are updates, plans, tasks, and assets. This LLM Wiki setup mainly uses tasks and assets. Updates and plans can be left for future team collaboration; no need to touch them in the first version.
Create a new project in WorkBuddy, e.g. llm-wiki; this becomes your cloud knowledge base workspace.
The "instructions" in a project are equivalent to AGENTS.md in the previous Obsidian setup; they determine what rules WorkBuddy follows when answering, compiling, querying, and writing back.
3. Write project instructions like this
When creating a project, put the following into "Instructions" and adjust it to your needs:
Be sure to select IMA in the skill options, because you'll later sync articles already saved in IMA. Experts and connectors can be skipped for now, unless your workflow really needs them.
4. Let WorkBuddy automatically create the directory structure
After the project instructions are written, there's no need to create directories manually. Type this in the project dialog:
WorkBuddy will read the project instructions and create raw/, wiki/, projects/, workflows/, index.md, log.md, etc. under the project asset root.
Check the asset directory, especially confirm whether raw/ and wiki/ were created successfully. All subsequent syncing and compilation depend on these two layers.
5. Sync IMA data with automation
In WorkBuddy, create an automation task that calls the IMA skill. You'll need your IMA API key here; just follow the prompts.
Write the automation task like this:
In the prompt, explicitly reference the IMA skill and clearly state the IMA knowledge base name and WorkBuddy project name.
For the first time, test manually. Confirm that WorkBuddy can read articles from the specified IMA knowledge base and create files under project assets /raw. After the test passes, enable daily automatic sync.
The first time you call the IMA skill, you need to pass in the IMA API key:
Sync result:
6. Compile new materials in raw
Once there's material in raw, you can do the first compilation.
Don't write something vague like "help me organize the knowledge base." Just do this:
Planning phase output:
After confirming the plan, let it execute:
After compilation is done, go to wiki/ under project assets and check the results. Focus on three things: whether there are source pages, whether there are concepts / topics / comparisons, and whether index.md and log.md were updated. If it only generated a few summaries, that's not enough — a real LLM Wiki needs to start forming structure, and pages should ideally have wikilinks.
7. Ask questions against the knowledge base
After that, you can ask questions directly in this project. For example:
I've verified that WorkBuddy's answer path follows the project instructions: first look at index.md, then find relevant wiki pages, and only go back to raw if that's not enough. The project instructions aren't just for show; they do constrain the answer path.
If the answer contains new insights, you can have it write them back:
Only with writeback does the knowledge base get richer the more you use it.
8. Use scheduled tasks to govern the knowledge base
If a knowledge base only grows and is never maintained, it will quickly become messy. Create another scheduled task for health checks:
Governance tasks should only give suggestions; don't auto-delete or auto-merge. A knowledge base is a long-term asset, so cleanup actions should be confirmed first.
9. Connect real work with projects
The knowledge base shouldn't stay only in raw and wiki. For example, if you want to write an article, create a project page directly under /projects in project assets:
From then on, writing articles, choosing topics, and organizing materials can all revolve around this project file. The value of this step is connecting the knowledge base to real tasks; otherwise it's just a database that can answer questions.
10. Which scenarios does it suit
It suits self-media best. Topic ideas, hot trends, competitor articles, reader feedback, and past viral posts go into IMA first, then sync to WorkBuddy's raw, and are regularly compiled into wiki. Create a project for each article, and after publishing, write back the title, read count, and comment feedback. Over time, it will understand your columns, topic boundaries, and headline preferences better and better.
It also works for daily office work. Create a project for each client or project. Meeting minutes, policy notices, WeChat Official Account articles, and project materials can all go into raw. Decisions go into wiki, action items go into log. Every morning have WorkBuddy read the project pages and yesterday's log to generate a morning brief; on Fridays, generate a weekly report. In these scenarios, the core value of AI is keeping work state uninterrupted.
11. A few things to note
This setup is essentially a cloud-based LLM Wiki.
WorkBuddy project assets are stored in the cloud. When processing, they're downloaded locally first, then uploaded after processing. With many files, the actual speed will be a bit slower, but the benefit is direct: you don't need to set up local Obsidian first; you can get running directly with WorkBuddy's project capabilities.
Project assets can be downloaded. After downloading, they can be imported directly into Obsidian, and they already contain bidirectional links. This means WorkBuddy + IMA can serve as a cloud solution first, and later migrate to Obsidian for local archiving and graph visualization. If you want a local notes solution with WorkBuddy + Obsidian, that's also entirely feasible — the underlying idea is the same; just replace WorkBuddy project assets with an Obsidian vault and project instructions with AGENTS.md.
I now divide the path into two options:
Finally
This experiment made me stop treating Obsidian as a mandatory entry point.
The real minimal LLM Wiki is about whether this pipeline can run:
IMA handles stable input of WeChat Official Account content. WorkBuddy handles projects, assets, compilation, query, writeback, and governance. That's enough to build a minimally viable personal LLM Wiki.
Obsidian still has value, but it's better suited as an enhancement layer. Get things running with WorkBuddy + IMA first, then decide whether to connect Obsidian — that's actually more stable.
I'm "Super Meng". I've long been tinkering with AI programming, agents, and automated workflows. I write practical notes on tools like Claude Code, Codex, and Hermes, and I also talk about how AI changes the real logic of development and content production.