Matt Wolfe's 15-Min AI Second Brain with Obsidian & Codex

Karpathy's Tweet Sparks the LLM Wiki Revolution

Matt Wolfe credits Andrej Karpathy's viral tweet (20.8M views) with igniting the 'LLM wiki' trend. Karpathy described experimenting with Obsidian markdown files in an IDE like Codex, where an LLM scans accumulated notes from YouTube transcripts and articles to auto-generate cross-links, mimicking Wikipedia's interconnected structure. "It's like your own little personal curated internet," Wolfe explains, emphasizing how the LLM uncovers hidden connections, such as linking AEO (Answer Engine Optimization) notes to Facebook ads under broader marketing themes. This 'mini internet' serves as context for queries, making users '100 times smarter' without manual organization.

Wolfe stresses this as the most popular non-coding use of Codex (OpenAI's code-focused IDE), distinct from ChatGPT's verbose style—Codex is tuned for concise, code-like responses. Hosts Kipp Bodnar and Kieran Flanagan highlight its appeal for non-developers like marketers and founders, positioning it as a productivity '10x unlock.'

Obsidian: Free Markdown Hub for Raw Knowledge Intake

At the core is Obsidian, a free markdown organizer that Wolfe calls 'just a text file with extra formatting.' Users download the app, create a 'vault' (folder), and it renders files as a navigable wiki with auto-generated indexes. Wolfe demos his vault: an index.md table of contents links to 'topics' and 'entities' pages, each aggregating sources like seven AEO entries. No manual filing—LLMs handle cross-linking.

Intake is seamless via the Obsidian Web Clipper Chrome extension. Wolfe clips a tweet or YouTube video (e.g., Future Tools' AEO episode), pulling full transcripts in one click to a 'RAW' inbox folder. "I literally just click this and it's now inside of my knowledge base," he says. This dumps unprocessed content, ready for LLM magic.

Codex Processing via Agents.md: Zero-Effort Organization

OpenAI Codex (free app, uses ChatGPT credits) opens the Obsidian vault folder directly. The 'agents.md' file acts as a master prompt, defining subprompts like 'ingest': read RAW files, dedupe URLs, generate/update topic pages, entity overviews, and syntheses. Wolfe commands: "Process all files in the raw folder," and Codex executes in ~2 minutes, moving files to structured folders.

"Agents.md literally tells Codex how to handle specific prompts," Wolfe notes. For example, it validates duplicates and links concepts. Free ChatGPT tiers work; paid plans ($20+/mo) scale usage. Plugins extend access: Gmail, Calendar, Drive, Slack, Notion, GitHub—even new Chrome integration for browser automation.

Overnight Automations and Proactive Outputs

Wolfe sets Codex cron jobs: a 12:50 AM automation scans RAW and processes overnight. "Your wiki builds itself while you sleep," he says. Daily 9 AM Slack brief proactively surfaces insights: "Based on recent saves, here's what to try next in your business." This shifts AI from reactive prompting to agentic proactivity—"the future is agents that come to you."

A demo query: "Best strategy for Future Tools in ChatGPT/Claude responses?" yields concise AEO advice (e.g., allow OI searchbot in robots.txt), citing wiki sources. Wolfe uses it for email triage: "Skim inbox, respond based on calendar + wiki."

Marketer Use Cases: From AEO Audits to Competitive Intel

Wolfe tailors for marketers: Codex + Chrome audits AEO (e.g., optimize as 'citable authority'), ad testing, product marketing. Competitive intelligence tracks sitemaps of OpenAI/Anthropic/Google, alerting to new posts instantly—"I see every new post the moment it ships."

Personal networking CRM: Log conference chats (name, topics), auto-dates, cross-references future interactions—"Never forget a conversation." Daily journaling links to entire knowledge base. Wolfe envisions team-scale: shared vaults, books as markdown.

"When we launched and nobody came, we realized X"-style stories emerge via resurfaced notes, turning consumption into action. 15-minute setup uses Karpathy's GitHub starter for self-building wikis.

"This is the single best AI productivity workflow I have ever seen," Bodnar declares. Free guide (link in description) provides steps.

Key Takeaways

• Download Obsidian (free), create vault, install Web Clipper for one-click YouTube transcripts/articles into RAW folder.
• Install Codex (free app), open vault folder, create agents.md with ingest prompt for auto-processing topics/entities.
• Set nightly cron: 'If unprocessed RAW files, process them'—builds wiki passively.
• Daily Slack automation: Proactive briefs from recent notes for business actions.
• Integrate plugins (Gmail/Calendar/Slack) for contextual tasks like email responses.
• Marketers: Query for AEO audits, ad tests, sitemap monitoring of competitors.
• Networking: Log convos as markdown; auto-links for recall.
• Start small: Free tier works; scale with paid OpenAI for heavy use.
• Avoid info hoarding—use proactivity to surface and act on knowledge.
• Future: Multiplayer team brains from single-player setups.