Source: raw/Build_Andrej_Karpathy_s_LLM_Knowledge_Base_for_Businesses_10x_Output.md — YouTube tutorial FAWm7DuFSPc by the “Dream Labs AI” channel (creator’s first-name not surfaced in transcript; new channel; uses the framing “the godfather of modern AI” for Karpathy). Walks through copying Karpathy’s GitHub gist + dropping it into a fresh Obsidian vault as CLAUDE.md and letting Claude Code scaffold the wiki — the exact pattern this karpathy/ project implements. 2026-10-27 refresh: adds a second one-prompt build without Obsidian, from Robo Nuggets (raw/Build_a_Self-Improving_Claude_Knowledge_Base_with_ONE_Prompt.md) — see § A no-Obsidian variant below.
A 10-minute meta-walkthrough on adopting the Karpathy LLM-wiki pattern specifically for a business context. Sources from Karpathy’s viral tweet + improved follow-up gist. The walkthrough collapses to: drop Karpathy’s idea file into a fresh Obsidian vault → Claude Code scaffolds the wiki structure → user drags raw assets (PDFs, transcripts, goals, competitor research) into raw/ → Claude Code compiles into the four-folder wiki (concepts / entities / sources / synthesis). Distinct from Karpathy’s own framing in that this is a business-applied walkthrough — a third-party operator’s tour of the pattern for SMB use.
Key Takeaways
The thesis
- Most users prompt LLMs with no business filter — the model returns “average IQ jumbled training data” outputs.
- Karpathy’s wiki pattern inserts a filter layer: prompts → business context (goals, voice, schema, history, competitors, rules) → LLM → re-filtered output → back into the wiki via a self-learning log.
- Compounding knowledge curve — wiki gets smarter with use rather than staying flat. Claims a “1,000-IQ employee that gets smarter every conversation.”
The anatomy diagram (creator’s framing)
- Far-left: all raw business data — PDFs, Google Drive exports, YouTube transcripts, Instagram reels, emails, messages, notes, competitor research, etc.
- Center: the schema (a
CLAUDE.mdrulebook for how to sort + handle that data). Borrows Karpathy’s rulebook by default; user can tweak. - Right: the Wikipedia-shaped output — Wikipedia-like cross-linked pages around your business, navigated visually in Obsidian.
The 10-minute walkthrough
- Install Obsidian (Mac or PC, free).
- Install Claude Code (terminal preferred; desktop app or other IDE work too).
- Open Karpathy’s idea file — the LLM-wiki gist on Karpathy’s GitHub. Click “raw” → ⌘A ⌘C to copy the ~91 lines.
- Create a new Obsidian vault — name it after the business (creator’s example: “Dreamlabs AI”); browse → Desktop → create. A default welcome page lands inside.
- Don’t delete the welcome page — Karpathy’s idea file has a transform function that uses it.
- Paste the 91 lines into the welcome page.
- Voice-prompt Claude Code: “I just pasted in an idea file from Andrej Karpathy. Can you please use the vault in Obsidian called [vault name] to build this out for me?”
- Claude Code scaffolds the structure:
raw/folder (with assets subfolder),wiki/folder withconcepts/entities/sources/synthesis/, plusCLAUDE.md,index.md,log.md. The welcome page is transformed into a user-guide. - Drag your raw business data into
raw/assets/— PDFs, transcripts, MDs, goals docs, screenshots — anything. No conversion needed. - Tell Claude Code “I’ve dropped my assets into raw. Please compile my business wiki.” Claude reads in parallel, writes source-summary pages, entity pages, concept pages, updates index + log. Reported time: ~8 min 42 sec for the demo business with 6 video transcripts + goals doc + competitor list + Alex Hormozi answers.
- Query it — “I’m planning a new YouTube video. Can you come up with an idea based on my business and past videos and write the intro?” Wiki-grounded output beats stateless chat.
Concrete asset types the creator dropped in
- Business goals doc — the “practical not novelty” rule for Dream Labs AI.
- YouTube transcripts — first 6 videos as
.mdfiles. - Articles — externally-written pieces “lighting the path” for the business direction.
- Project files — YouTube channel, paid community.
- Interviews — Claude-conducted self-interview using Alex Hormozi’s 12 questions as the question set.
- Competitor research — files on named competitors (Greg Eisenberg called out as a Dream Labs competitor).
Role-model “skills plugin” concept (creator’s own work-in-progress)
- Take all of a public mentor’s publicly-available data (Tony Robbins cited as example).
- Build a “skills file” — described as “like a Nintendo gaming cartridge that you plug into your wiki.”
- When you query the wiki, the role-model mentor’s perspective guides the next-step suggestions and output framing.
- This is the same shape as the operator-curated skill catalog but oriented toward expert-emulation rather than task-automation.
Gap-detection query
- One of the highest-value queries the creator demonstrates: “Tell me where the gaps in my knowledge are for my business.”
- Returns specific holes in the wiki that the operator can fill — the wiki audits itself for completeness.
A no-Obsidian variant: Robo Nuggets’ one-prompt knowledge base (2026-06)
Creator: Robo Nuggets (Jay) · URL: https://www.youtube.com/watch?v=K2BpNt3UBOQ · Duration: 12:35 · Platform: YouTube · Published: 2026-06-08
Same pattern, fewer tools: no Obsidian, no vector or RAG database, just Claude Code (he demos in the Claude desktop app’s Code tab) and one setup prompt that embeds Karpathy’s LLM-wiki pattern. Claude states its plan, asks for one confirmation, then creates the folders. Jay cites Karpathy’s own knowledge base as about 100 articles and 400,000 words (his figure).
Five building blocks:
| Block | What it is | Equivalent in this vault |
|---|---|---|
raw/ | Unorganized inputs, messy by design; Claude’s deep-research results land here too | raw/ + ai-research/ |
wiki/ | Organized, linked summaries plus index (table of contents of concepts, entities, sources), log (dated adds and removals) and processed (which raw files are already ingested) | topic folders, _master-index.md, log.md, .manifest.json |
outputs/ | PDFs, decks and apps generated from the wiki | output/ |
CLAUDE.md | The operating manual, maintained by the agent itself | vault CLAUDE.md + operations/ |
| Dream sequence | A scheduled routine (weekly by default) that ingests anything new in raw/, then lints for contradictions, stale claims, duplicates and orphans; editable under Routines in the desktop app | bin/post-ingest + Lint, run on demand |
Using it.
- The setup prompt ends with a refinement question (what the wiki is for) and asks whether to research the topic from the web or let you supply files.
- Save from your phone: share an article link to Claude and say “save this”; because of the
CLAUDE.mdrules, it lands inraw/and the wiki updates. - Change the cadence by asking. His tweak: stop auto-updating the wiki on every save, let the dream sequence handle ingest, and run it daily. Claude edits its own
CLAUDE.mdand the routine. - The dream sequence can be run manually: Claude lists
raw/, compares it against theprocessedregistry and folds only the new files in.
What this vault adds beyond it: per-topic folders past the 4-folder default, sha256-hashed manifest entries instead of a plain processed list, QMD hybrid retrieval, and contradiction callouts tracked in frontmatter. ^[inferred] Jay later describes the same plain-text, no-embeddings approach for a 420,000-file workspace in ARMS Agentic OS.
Validation against this vault
This walkthrough is essentially a guided tour of the same machine you (the reader) are interacting with right now. Compare specifics:
| Walkthrough’s structure | This vault’s structure |
|---|---|
raw/assets/ for source files | raw/ (with no nested assets/) |
wiki/concepts/ entities/ sources/ synthesis/ | wiki/{topic}/ topic folders (deliberately divergent — this vault’s 20-topic schema is more granular than Karpathy’s 4-folder default) |
CLAUDE.md as schema | CLAUDE.md as vault schema (full Operations spec) |
index.md master TOC | _master-index |
log.md for self-learning loop | [[log |
| Welcome page transform | (this vault skips welcome transform) |
The pattern is the same; this vault diverged on folder layout to support multi-topic navigation at the 200+ article scale (see CLAUDE.md § Scale Awareness).
Open Questions
- Creator’s identity — “Dream Labs AI” channel; first name not surfaced in transcript. Creator references Alex Hormozi’s 12 questions + Greg Eisenberg as competitor.
- Karpathy’s improved gist URL — referenced as the “updated” version of the viral tweet, linked in video description. Worth pulling as a primary source comparison vs. the original tweet and this vault’s
CLAUDE.md. - The “21 million views” stat — claimed for Karpathy’s original tweet; not independently verified in transcript.
- Role-model skills-plugin format — described as gaming-cartridge-shaped, creator says they’re “personally working on” it. No repo or shipped artifact yet.
- Self-learning log mechanism — creator claims wiki “updates its own memory” with each query feedback. The exact write-back loop is unspecified beyond “log.md tracks every interaction” — likely the same Log operation this vault implements (see
CLAUDE.md§ Log).
Related
- _index — topic root for community implementations of this pattern
- pipeline-diagram — visual of the ingest pipeline
- joshpocock-vault — another community implementation (joshpocock “Stride starter”)
- synthadoc — axoviq-ai/synthadoc implementation
- matt-wolfe-ai-second-brain — another third-party YouTube build (Matt Wolfe), adding journal + CRM layers and an hourly Codex automation loop on top of the wiki
- from-vibe-coding-to-agentic-engineering — Karpathy’s own talk on the agentic engineering shift
- ben-five-skill-aios-setup — Ben’s 5-skill AIOS bundle (the
/os-optimizerskill explicitly cites the Karpathy LLM-wiki as one of its frameworks) - nate-herk-claude-code-operating-systems-course — Nate Herk’s AIOS course (parallel framing — Three Ms + Four Cs vs Karpathy’s raw → schema → wiki)
- seven-claude-skills-run-my-business — operator-curated skill catalog; closest existing article to the role-model-plugin concept
- ARMS Agentic OS (Robo Nuggets) — the same creator’s full workspace setup, where this knowledge base becomes the memory layer
Try It
- You already have this — this vault IS a working instance. The deltas vs the walkthrough are documented above.
- For someone starting fresh: copy Karpathy’s gist into a new Obsidian vault’s welcome page, open Claude Code in that folder, run the voice-prompt from step 7 above, then drop your business raw files in.
- If you have an existing Karpathy wiki: test the gap-detection query (
"Tell me where the gaps in my knowledge are for my business") — same effect as a query-log audit but expressed conversationally. - Adopt the Alex Hormozi 12 questions — useful seed for
raw/if the business doesn’t have rich existing documentation. - Investigate the role-model-skills-plugin idea — package a public mentor’s body of work as a skill folder under
.claude/skills/<mentor-name>/with aSKILL.mdthat orients answer framing in that mentor’s voice. Adjacent to existing skill patterns (skills are folders). - Try the dream-sequence split: keep ingest-on-save off and schedule one daily routine that ingests new
raw/files and then lints, so saving a link never triggers a full compile mid-task.