how to set it up:
1. clone the course repos you want into raw/ (all free on https://t.co/auzn6G9haI)
2. tell Claude Code: read everything in raw/, write one wiki page per concept in wiki/, link each page to the chapters that cover it
3. add a CLAUDE.md: raw/ is read-only, answer from wiki/ first, always cite the chapter path
4. ask it anything
the 9 repos I used: course, agents-course, smol-course, mcp-course, cookbook, deep-rl-class, audio-transformers-course, computer-vision-course, diffusion-models-class
the idea comes from Andrej Karpathy's LLM wiki
I FED CLAUDE EVERY FREE HUGGING FACE COURSE. NOW IT ANSWERS WITH THE CHAPTER.
9 course repos from GitHub, 866 files in English, 7,955 sections. Karpathy's "second brain" idea, pointed at learning instead of notes
the setup:
→ raw/ holds the 9 repos untouched: LLM course, agents, smol, MCP, cookbook, deep RL, audio, vision, diffusion
→ Claude wrote one wiki page per concept, 40 in total, each linking to the chapters that cover it
→ CLAUDE.md has 5 rules. the one that matters: always cite the chapter path
→ 2,764 links between concepts and chapters
then I asked it "where do I learn MCP from zero?"
it read wiki/mcp.md, found 51 files that cover MCP (47 in the MCP course, 4 in the agents course) and gave me an order: intro, key concepts, clients, tiny agents, then build your own server
no tutorial hunting, no 40 open tabs. one question, one reading list with real file paths
how to set it up yourself ↓
@mojeskoqq curious how it holds up after a few hundred sources. does the CLAUDE.md need rules for when to merge pages, or does Claude handle that on its own?
@0x_fokki the part that got me: it rendered a still and fixed the gaps for 3 rounds before touching the film. that's exactly what most people skip. how long did the whole thing take?
the 4 prompts (paste one per new chat):
AI AGENT:
A highly detailed 3D animated feature film character design of an AI AGENT, a fully modeled 3D character that looks ready to drop into a video game or animated movie. A small friendly robot with a large round helmet-like head, a glossy black face screen with big glowing blue pixel eyes, a short antenna, scuffed off-white armor panels with visible wear and dents, standing confidently with hands on hips. Beside it: a long paper scroll checklist unrolling onto the floor with exactly 12 lines, numbered 1 to 12 in correct order, each with an empty checkbox and a blank line, numbers clean and readable with nothing skipped or repeated; on the other side a worn retro computer terminal showing a ">_" prompt, with a glowing blue holographic loading ring floating above it. Realistic material wear, subsurface shading, no logos anywhere, rendered in a modern 3D engine, cinematic 3D rendering, soft studio lighting with gentle rim light, full-body composition, clean white background, vertical 4:5
PROMPT ENGINEER:
A highly detailed 3D animated feature film character design of a PROMPT ENGINEER, a fully modeled 3D character that looks ready to drop into a video game or animated movie. A thoughtful young man with curly brown hair, round wire glasses and light stubble, a yellow pencil tucked behind his ear, one hand on his chin and a knowing half-smile, wearing a chunky dark green knit cardigan with a Fair Isle pattern over a cream button-up shirt, rolled-up brown corduroy trousers with a leather belt, worn brown leather lace-up boots. At his feet: a large unrolled parchment scroll with the handwritten words "You are an expert..." clearly legible, a brass magnifying glass, and a stack of old leather-bound books with a yellow sticky note on top that reads "be concise". All text spelled exactly as written. Rich fabric fibers, realistic material wear, detailed cloth folds, subsurface skin shading, no logos anywhere, rendered in a modern 3D engine, cinematic 3D rendering, soft studio lighting with gentle rim light, full-body composition, clean white background, vertical 4:5
VIBE CODER:
A highly detailed 3D animated feature film character design of a VIBE CODER, a fully modeled 3D character that looks ready to drop into a video game or animated movie. A relaxed young developer with messy dark curly hair, light freckles, sleepy half-closed eyes and a smug little smile, hands in the pockets of an oversized charcoal hoodie, over-ear headphones resting around his neck, baggy olive cargo pants, chunky plain cream sneakers. Next to him on the floor: an open laptop showing a code editor on one side and a photo gallery preview on the other, three crumpled paper coffee cups (one tipped over) and a yellow rubber duck. Rich fabric fibers, realistic material wear, detailed cloth folds, subsurface skin shading, plain unbranded sneakers and laptop, no logos, no brand marks, no stripes or symbols anywhere, rendered in a modern 3D engine, cinematic 3D rendering, soft studio lighting with gentle rim light, full-body composition, clean white background, vertical 4:5
STARTUP FOUNDER:
A highly detailed 3D animated feature film character design of a STARTUP FOUNDER, a fully modeled 3D character that looks ready to drop into a video game or animated movie. A confident young man with tousled dark brown hair and light stubble, gazing upward with a visionary, slightly dreamy look, wearing a black ribbed turtleneck sweater, slim dark grey jeans and plain white sneakers. He holds a smartphone in one hand and hugs a printed pitch deck to his chest with the word "AI-powered" on the cover above a small mountain photo. Behind him: a whiteboard on an easel with a hand-drawn bar chart and an arrow shooting up and to the right. To the side: a small round wooden side table with a potted green plant, a plain silver energy drink can, a spiral notebook and a pen. All text spelled exactly as written. Rich fabric fibers, realistic material wear, detailed cloth folds, subsurface skin shading, plain unbranded phone, can and sneakers, no logos anywhere, rendered in a modern 3D engine, cinematic 3D rendering, soft studio lighting with gentle rim light, full-body composition, clean white background, vertical 4:5
I PAID FOR TWO IMAGE MODELS. THE FREE ONE WON.
same 4 prompts, 4 models, first attempt only. no rerolls
the test: 3D animated characters for tech jobs, with text that has to come out right. a scroll that says "You are an expert...", a sticky note that says "be concise", and a robot next to a checklist numbered 1 to 12
→ GPT (paid): 4/4 clean. every word, every number
→ Grok (paid): broke the checklist. 10, 10, 11, 12, 12. then a second list that skips 3, 7, 9 and 11
→ Gemini (free): broke it too. 9, 9 ... 12, 12. plus fake scribbled text under "You are an expert..."
→ Copilot (free): 1 to 12 in order, every word right, and the most "animated movie" look of the four. it did add the job titles on its own
counting to 12 turned out to be the real test. half the models couldn't do it
my pick for this style: Copilot. free, and on par with paid GPT
all 4 prompts in the reply ↓
@steipete The rule itself might be reasonable, but shipping it silently with no toggle is what makes people nervous. Has anyone seen it documented anywhere?
@PJaccetturo $28K and 62 million views is a better return than most studio films 😂 What took the biggest chunk of the budget, the generations or the edit?
@fluid_ai_ Native sound is the part that got me. A world that hears itself feels way more alive than one that just looks right 🌊 Did you find anything on how long the memory actually lasts before it forgets?
@AIBoticssq Feels more like a correction than a pop to me. The compute is real, the open question is how long investors wait for revenue to catch up 📉 What timeline would you give it?