Matt Wolfe builds a personal knowledge system with Obsidian and Codex in this May 2026 walkthrough. Saved sources become a linked wiki that supports questions, journal responses and contact lookup. He adapts Andrej Karpathy’s LLM Wiki concept, adding his own workflows around it. The Hostinger segment at 06:14–08:03 is sponsored and separate from the build.

Give saved material a reusable structure (01:07–06:12)

  • Wolfe puts a wiki at the center of three connected collections: reference material, contacts and journal entries. Other users could substitute research papers, classroom notes or client records.
  • The proposed processing step summarizes sources, extracts people and concepts, and links related notes. A tool mentioned in several sources can have a page connecting those appearances.
  • Journal responses should draw on that saved material and earlier conversations. Wolfe wants relevant ideas resurfaced when he needs them, rather than left in a folder he rarely opens.

Keep sources separate from generated pages (08:03–12:54)

  • Obsidian provides the reading interface for ordinary Markdown files. Codex opens the same vault folder as its project and edits those files directly.
  • The first generated scaffold contains 51 files. Wolfe asks Codex to remove the extras and return to the smaller architecture he requested.
  • The resulting structure separates raw/, containing source material, from wiki/, containing generated pages. AGENTS.md holds operating instructions, index.md catalogs pages, and log.md records changes.
  • Keeping the source layer available gives readers somewhere to check the generated synthesis. The later workflow changes source metadata and file locations, so “immutable” here needs a precise definition.

Check ingestion rules against actual files (12:56–18:09)

  • Wolfe configures Obsidian Web Clipper to save into raw/. His template includes the source title, URL and capture date. He explicitly distinguishes that date from original publication.
  • In his demonstration, the clipper imports YouTube transcripts. Saving alone does not build the wiki; he initially tells Codex to process the folder manually.
  • Processing preserves the original material and creates linked source and concept pages. Additional imports expand the index and graph.
  • Channel attribution needs a correction. His first instruction adds the channel name to generated wiki pages, while he wanted it on the original source note.

Questions can become new wiki pages (18:13–21:23)

  • Wolfe asks for advice about starting a difficult task. Codex checks the index, reads relevant pages and produces an answer grounded in saved material.
  • That interaction also updates the index and log, and creates a reusable page linked to its sources. The answer becomes part of the maintained collection.
  • Wolfe edits AGENTS.md to move ingested files into raw/processed/. He also requires generated pages to link back to source notes, reducing orphaned summaries.

Add contacts and journals through explicit instructions (21:25–28:22)

  • A CRM, or contact-record system, uses one file per person plus an alphabetical index. Wolfe tests adding a record and later asking where they met.
  • Journal mode saves the entry and subsequent conversation, updates an index, and records a summary in the log. Its instructions allow responses to draw on the wiki, past entries and contacts.
  • The sample journal response uses relevant creator notes from the wiki. It also uses the model’s general knowledge, so readers should distinguish cited source material from additional advice.

Schedule processing and preserve a backup (28:24–31:39)

  • After reprocessing earlier imports under the revised rules, Wolfe configures an hourly Codex automation to check for unprocessed files.
  • He selects the existing local project so the automation works on the vault itself. The model and reasoning setting shown are May 2026 choices, not permanent requirements.
  • He creates a private GitHub repository, pushes the vault, and extends the automation to commit and push after processing. Personal journals and contact details make that privacy setting consequential.

Caveat: the demonstrations show a small working vault, not reliable citation accuracy at scale; local Markdown storage also does not make hosted model processing private.