Part 6 – I Built the Next Layer. It’s a Research Wiki That Grows Itself.
I thought the next layer would be more Obsidian configuration. It wasn’t.
More graph view, more backlinks, more manual connection-making. That is what I used to do in Roam. I would spend real time threading notes together, following paths between ideas, maintaining a map of how everything connected.
I did not do any of that. I built a wiki instead, and Claude Cowork maintains it.
When I was getting my doctorate, I uploaded everything to Zotero. Tagged it, linked it, summarized it by hand. It took real time, and what I got back was an organized pile. The pile did not think. It did not notice when two papers contradicted each other or tell me what a new source meant for something I had already read. I was doing all of that work myself, in my head, every time.
I thought I was being clever when I started using Atlas.ti. It has this visual network view where you can map relationships between ideas, draw connections between quotes, build a picture of how everything fits together. I spent real time in there during my doctorate. The network looked impressive. But I was still the one making every connection. I decided what linked to what. I noticed the contradictions. The tool rendered what I built. It did not think.
I tried Elicit. I tried Research Rabbit. Both are good at finding papers. That is where they stop. Consensus was different because it tells you what the papers actually say, not just that they exist. And once I paired it with Claude Cowork and the ingest skill, the whole thing clicked. The literature review runs first. The output drops into raw-sources. The skill reads it and threads it into everything the wiki already knows. The pile became a system.
Here is how it works.
I have two sources of incoming research. The first is Consensus, which I use to run preliminary literature reviews on topics I am actively working on. I give it a research question. It surfaces peer-reviewed papers, synthesizes findings, and produces a document. That document goes into a folder called raw-sources. The second source is anything interesting I find while reading — articles, whitepapers, newsletter pieces, practitioner reports. Those go into raw-sources too.
From there, I run a wiki-ingest skill. It reads the new source, writes or updates a topic page in my wiki, cross-links to related pages that already exist, flags any contradictions with what the wiki already contains, and appends an entry to a running log. I do not write any of this. I drop a file and run the skill.
The wiki lives in the same plain text folder as everything else. Obsidian renders it. The graph view, which I mentioned but had not yet populated in Part 2, now shows real connections. Not connections I built by hand. Connections the ingest skill wrote, in wikilink format, across every page it has ever touched.
The tool renders. The AI thinks. That sentence from Part 2 turns out to apply here too.
The thing that surprised me is what happens over time.
Each new source does not just add a page. It enters a system that already knows what I think about adjacent topics. When the ingest skill reads a new paper on AI and first-generation students, it checks what the wiki already has on custom GPTs, cognitive debt, and equity research. If the new paper contradicts something the wiki already contains, it flags the contradiction. If it confirms something, it adds the citation to the existing page. The knowledge base does not just grow. It develops opinions.
I have 17 pages covering roughly 200 papers right now. When I ask a question against the wiki, the answer is not drawn from a single document. It is drawn from the accumulated record, cross-referenced and contradiction-checked, in a way that no individual session could reproduce from scratch.
Before this, every research project started over. I found sources, read them, and retained whatever I could hold in my head until the project was done. Then the next project began the same way. The reading did not compound.
Now it does. Every source I ingest makes the next question easier to answer. The wiki gets more useful the longer I use it.
That is what the next layer turned out to be. Not more configuration. Not more backlinks by hand. A system that builds its own connections and keeps track of its own contradictions, while I focus on the questions worth asking.
