Frequently Asked Questions About Matt Giaro AI Second Brain Build
23 answers covering everything from basics to advanced usage.
// Basics
What exactly is a 'second brain' in the context of AI?
A second brain is an external system that stores and organises your knowledge so you don't have to hold it all in your head. The AI version adds an active layer: instead of just archiving notes, an AI reads your saved content, builds a cross-linked wiki from it, and resurfaces relevant pieces when you query or journal — turning passive storage into an interactive knowledge partner.
What is the Dumping Ground Problem?
It's Matt Giaro's term for the failure mode of most second brain systems: information gets saved but never resurfaces, making the system a passive archive instead of an active tool. This framework solves it by requiring a processing pipeline — every source is wiki-fied, cross-linked, and made queryable so knowledge comes back out when you need it, not just when you go looking.
Why is the wiki considered the centre rather than one of three equal pillars?
The wiki is the gravitational core because the Journal and CRM derive their value only by being grounded in and cross-linked to it. Remove the centre and the other layers become isolated silos. A journal entry is valuable because it references your saved knowledge; a CRM contact is valuable because it connects to the ideas you discussed. The wiki is what makes everything else meaningful.
What is Karpathy's LLM Wiki and why does this framework use it?
Andrej Karpathy's LLM Wiki is an architectural blueprint for using an LLM to maintain a self-building, cross-linked markdown wiki from raw source material. Giaro uses it as the foundational wiki layer because it establishes a proven structure for entity extraction and cross-linking, then extends it with Journal and CRM pillars. A useful first seed is ingesting the Karpathy LLM Wiki GitHub page itself.
// How To
How do I configure the Obsidian Web Clipper correctly?
In the Web Clipper settings, add your vault name exactly as it appears in Obsidian, set the default template to pull source title, source URL, creation date, and a 'web-clip' tag, and set Note Location to 'RAW'. This ensures all clipped content lands in the correct folder. The Web Clipper automatically pulls full YouTube transcripts, making it the primary ingestion mechanism.
How do I set up the hourly automation to process new content?
In Codex, go to Automations → New Automation. Title it 'Process Second Brain RAW Files', set work tree to Local, select your Second Brain project, choose an hourly schedule, and use the prompt: 'If there are any unprocessed files inside the RAW directory, please process them.' Use the strongest available model on high reasoning. After setup, you only need to clip content — processing happens automatically.
How do I add a Journal pillar to my system?
Prompt Codex to update agents.md with journal rules: if a chat begins with 'journal', save the full conversation as a markdown file in /Journal named [date]-[short-title].md, add an entry to /Journal/index.md, log it in log.md, and ground the response in wiki content, past journal entries, and CRM data. The 'journal' trigger word is essential — without it, entries are treated as wiki queries and never saved.
How do I add someone to my CRM?
In Codex, say 'Add to CRM: [Name] — met at [event] in [year], discussed [topic], their role is [X], follow up about [Y].' The system creates or updates a file in /CRM named after the person, capturing contact details, how you met, relationship context, and notes. It maintains /CRM/index.md alphabetically with a short bio per contact and logs the update in log.md.
How do I connect a GitHub repo for automated backup?
Create a new private repository on GitHub, then in Codex prompt: 'Commit this current version to my private GitHub repo [URL].' Then edit your hourly automation prompt to append: 'Once everything is processed, commit and push the current version to the main branch on GitHub.' Backup now happens every hour alongside processing, with no manual effort.
// Troubleshooting
My wiki pages are orphaned and the graph view looks flat — what's wrong?
Your agents.md is likely missing the cross-linking instruction. Without the rule to 'cross-link generated wiki pages back to the original source page' and extract entities, pages never interconnect and the graph stays disconnected. Edit agents.md to enforce entity extraction (people, companies, tools, ideas, themes) and bidirectional cross-linking, then reprocess your sources.
My journal entries aren't being saved as files — why?
You're probably not starting the chat with the 'journal' trigger word. The system only routes to journal-handling logic if the chat begins with 'journal' on its own line. Without it, the entry is treated as a wiki query and no journal file is created. Always lead with the trigger, then write your entry on the following lines.
Codex built dozens of extra files when setting up the architecture — how do I fix it?
Codex tends to over-build if not constrained. Explicitly prompt it to prune back to the minimal Karpathy structure: /RAW, /RAW/Processed, /RAW/Assets (optional), /Wiki, agents.md, index.md, and log.md. Tell it 'Remove everything except these elements and rebuild only this minimal structure.' Anything beyond these core files adds noise without value early on.
My wiki pages are shallow with poor entity extraction — what's causing it?
You're likely using a weak AI model for processing. The hourly automation and manual processing should use the strongest available model on high reasoning. Lightweight models produce shallow wiki pages with poor entity extraction and weak cross-linking. Upgrade the model in your automation settings and reprocess the affected sources to rebuild richer, better-connected pages.
The YouTube channel name is showing on my wiki pages instead of the source — is that right?
No. The channel name belongs in the front matter of the original source file in RAW, not on the generated wiki page. Update your agents.md processing rules to specify that for YouTube sources, the AI extracts and adds the channel name to the original source page front matter. This keeps wiki pages clean and attributes provenance correctly.
// Comparisons
How does this compare to using ChatGPT's memory feature?
ChatGPT's memory stores scattered facts it chooses to remember and lives inside one chat interface. This framework builds a structured, inspectable knowledge base you own locally in Obsidian, with explicit cross-linking, a graph view, version history, and grounding in your deliberately saved content. You control exactly what's ingested and how it's processed, and it works across any AI agent rather than being locked to one vendor.
How is this different from a standard RAG (retrieval-augmented generation) setup?
A typical RAG setup embeds documents in a vector database and retrieves chunks by similarity. This framework instead builds a human-readable, cross-linked markdown wiki that both you and the AI can read directly. It emphasises entity extraction, bidirectional linking, and a visual graph view, making the knowledge structure transparent and editable — not a black-box embedding store you can't inspect.
How does the AI second brain compare to Notion AI?
Notion AI operates on your Notion workspace and is tied to that platform and its pricing. This framework is local-first in Obsidian with plain markdown files you fully own, driven by a single agents.md constitution you can edit freely. It also adds automated hourly processing, a self-building wiki, and the ability to swap between AI agents like Codex or Claude Code without migrating your data.
// Advanced
Can I replace the CRM pillar with something domain-specific?
Yes — the CRM is fully customisable to your domain. Students can replace it with a Classroom Notes pillar, fitness enthusiasts with workout logs, consultants with client records, and cooks with recipes. Only the Wiki/Knowledge Base is non-negotiable. You define the relational or logging layer that fits your needs and configure its behaviour in agents.md, keeping it cross-linked to the central wiki.
How do I iterate on system behaviour after the initial build?
Everything is prompt-driven, so any behavioural change is made by editing agents.md directly in Obsidian or prompting Codex to update it. Add new folder categories, additional entity types to extract, or stricter cross-linking rules there. Use the Obsidian graph view to monitor how interconnected the wiki is becoming — a densely linked graph is the best signal of system health; a flat one signals broken rules.
How much content should I seed the system with before it's useful?
Seed at least 5–10 sources to give the wiki meaningful cross-linking from the start. With too little content, entities won't have enough to connect to and the graph view stays sparse. A strong first seed is the Karpathy LLM Wiki GitHub page itself, followed by a batch of articles and YouTube videos in your core domain so related concepts begin interlinking immediately.
What should be inside the agents.md processing pipeline?
The ingestion pipeline should: (1) read the source from RAW, (2) create or update wiki pages, (3) update entity/concept/topic/overview/synthesis/comparison pages, (4) cross-link generated pages back to the original source page, (5) update index.md, (6) append an entry to log.md, and (7) move the source from RAW to RAW/Processed. For YouTube, add the channel name to the source page front matter.
How long does processing take per batch of content?
Processing time scales with content volume — allow roughly 3–6 minutes per batch. In Codex, open a new chat in your Second Brain project and prompt 'Process the files inside the RAW folder.' The AI reads each source, generates wiki pages, extracts entities, cross-links related pages, updates index.md and log.md, and moves files to RAW/Processed. Larger or denser sources like long transcripts take longer.
Is the graph view just cosmetic or does it serve a real purpose?
It's a genuine diagnostic tool, not decoration. The graph view is the primary signal of whether the system is working. A densely interconnected graph indicates cross-linking and entity extraction are functioning correctly. A flat, disconnected graph after weeks of use means the cross-linking rules in agents.md are failing — a cue to inspect and fix those rules before adding more content.