An LLM wiki in Obsidian
I went looking for structure and came back with two hygiene rules. My layers were already finer than the video's, and the thing I wanted most I put on hold.
I went for the structure and came away with two hygiene rules. A 25-minute video shows a published structure — the LLM wiki — implemented in Obsidian with Claude Code. Source material, wiki and rules in three layers, building a company brain the model reads and organises itself. Held against mine, my layers were already finer, and my nightly check was already running with a human approval gate on top. What actually remained was not structure but two operating rules, and the one I wanted most I put on hold.
What it is
A case study of actually running a published structure. The core proposition is one line on a slide — “don’t re-enter knowledge each time; compile it once and keep accumulating.” Instead of handing over material again and again, you attach the model to a hub you built once.
He calls it a library and divides the space into five. Workshop / library / personal vault / conversation channels / manual archive. Work happens in the workshop, memory lives in the library, personal things in the vault — and only reviewed, finalised material moves into the library. It has been running two or three months.
Five operating principles.
| Rule | What it means |
|---|---|
| Finalised only | Work in progress never enters the library |
| One item, exactly one place | No duplicates |
| Unsure → inbox | Defer the judgement, accept it now → auto-sorted overnight |
| Move instead of delete | Old things go to legacy, not to the bin |
| Log the access | Which model looked at what, and when, goes into a log file |
A hygiene check runs at 3:15am and a weekly report lands on Monday. But the supervision is human — the automatic check proposes improvements and a person reads them.
The most practical technique was the recording
It generates work manuals automatically from screen recordings. A browser extension he wrote records a task while capturing clicks, screenshots and narration, then turns it into a markdown manual. The output is one document plus an image folder, and handing that folder over with “add this to the wiki” means the original is preserved in the source layer and a separate derived page appears.
The use case is clear — onboarding, delegation, handover. From then on, a request like “summarise monthly revenue” comes back with wiki context and sources folded in.
What broke
The structure isn’t new, and the video doesn’t hide that. The three layers come straight from a published concept, credited up front. The value here is not the concept but two months of operating notes.
There are no verifiable numbers, though. Nothing measures how much better things got with the wiki attached, or how much hallucination dropped. “Fewer hallucinations, higher accuracy” is an opening-slide claim with nothing behind it.
And the back half is a product pitch. He launches his own app and points at a course — not a flaw, but you have to read the operating-notes part and the promotional part separately.
Held against my own setup
My layers were already split finer. Where the video splits into three I use four, and my nightly check runs with a human approval gate behind the automatic proposal. The graph visualisation was already built too.
So what remained wasn’t structure, it was two rules.
One — one item, exactly one place. My vault already has a recorded problem: the same content stubbed out in several places. “When unsure, put it in the inbox and sort later” prevents the problem from occurring at all — it defers the judgement without multiplying the locations.
Two — the access log. Right now my main session, my scheduled runs and my messaging worker share one config folder. Nothing records who read or changed what, so when something drifts, finding the culprit costs time. One log file covers it.
Three — the screen-recording manual is on hold. It was the thing I wanted most, and auto-documenting QA reproduction steps with clicks and screenshots sounded immediately usable. But this is an item I already evaluated once and deferred. The reason not to push it still holds — nothing is actually blocked today by reproduction steps going unwritten. With a friction sample of zero, bringing in a tool leaves you with only the cost of maintaining it.
Verdict
| What | Verdict |
|---|---|
| One item one place · unsure → inbox | adopt. Prevents duplicates instead of catching them afterwards |
| Model access log | adopt. Needed now that several paths share one config |
| Screen recording → manual | still on hold. Friction sample of zero Reopens: when something is actually blocked by an unwritten reproduction step |
| Three-layer structure | already have it. Split into four |
| Nightly check · graph view · hub | already have it |
| Citing its numbers | ⛔ don’t. Nothing here was measured |
The lesson I paid for here: where someone is ahead of you and where they are behind sit inside the same material. My structure was ahead; my operating rules were behind. Look only at the structure, close it with “I have all this already”, and both rules are gone.
And ⚠ the thing you want most is not the thing you need most. The screen-recording manual was the most impressive part of this video, and my friction sample was zero. Fail to separate impression from need and the tools simply multiply.