Start by creating a single project folder with three subfolders—raw/, wiki/, and outputs/. The raw/ folder holds everything you’ve collected: pasted articles, screenshots, meeting notes, research papers—no organization or renaming. The wiki/ folder is where AI turns that heap into structured content. Outputs/ stores answers you’ve asked the AI to generate. No databases or fancy apps—just text files and folders, mirroring how @karpathy does it. Next, automate collection and organization. Use agent-browser, a free CLI tool from Vercel Labs, to have your AI scrape web pages—including JavaScript-heavy sites or pages behind logins—and dump them into raw/. Write a simple schema file (called CLAUDE.md or AGENTS.md) at the project root. That file tells the AI how to structure your wiki. Then point Claude Code (or a similar AI-coding tool) at your folders. It reads raw/, applies the schema, and writes an INDEX.md plus one markdown file per topic inside wiki/, linking related pages and summarizing sources. Once the wiki grows beyond ten articles, start asking questions like “What gaps remain in my understanding of X?” or “Compare source A and source B on concept Y.” Save those AI answers back into outputs/ or have the AI update the relevant wiki pages. Every edit makes the system smarter—until errors compound. Avoid that by running a monthly health check: have the AI flag contradictions, unbacked claims, and missing topics, then suggest new articles to fill gaps. No need for Obsidian or 47 plugins. Flat files, one schema, and a scraper let your knowledge base scale without tool bloat.