Loading video...
Video Failed to Load
Used Karpathy's LLM Wiki to turn Claude Code into a self-evolving system. Fed it my frontend & CRM sources, and it built a fully updating dashboard with Shadcn packages and constantly improving and linking everything for smarter outputs! Full setup video:
133,755 views • 7 months ago •via X (Twitter)
22 Comments

I built an open-source Claude Code skill that implements @karpathy's LLM-Wiki concept. It's not a second brain, it's a living vault. Every source you feed it rewrites existing pages, resolves contradictions, and synthesizes patterns you didn't ask for. The vault doesn't grow. It evolves. 22 commands, scheduled agents, thinking tools that argue with you using your own notes.

The self-evolving part is the breakthrough. Most AI setups are stateless, you start over every conversation. Once the system compounds knowledge across sessions, everything changes.

Where's your GitHub link 👀 🤔

the risk with self-referential codebases is you optimize for what the LLM thinks is improving rather than what actually ships. karpathy's wiki concept is solid but feeding it your own generated code can compound errors fast

You turned Claude into a self-updating dev? Respect. Drop the vid and repo, imma feed it my dumpster CRM and watch it glow up lol

Did you try with OpenEvolve too?

the self-evolving part is what matters. I do something similar but with persistent memory files instead of a wiki. every session writes observations about what worked and what didnt. next session reads them on boot. after a few weeks the system knows things about the codebase that no single session could discover

100% on statelessness. Claude Code and Cursor don't lack intelligence — they lack persistent memory between sessions. I solved this for my daily NASDAQ bot by making the repository the source of truth (NOTES.md + auto-push WIP checkpoints). Session crashed? Just send "." and the agent inspects memory and resumes without asking. Demo in the video above!

Karpathy's wiki as the foundation is a smart move - that man just quietly ships the best learning resources. 👀

does it stay accurate after a few update cycles? every self-evolving setup i've tried eventually starts confidently rewriting correct stuff

The interesting shift is using the knowledge base as shared context instead of a one off prompt. Once the source docs stay fresh, the agent output gets much more consistent and reviewable.

the self-evolving part is what makes this interesting. most people use claude code as a one-shot tool. feeding it your own data sources and letting it iterate on its own context is where it gets weirdly powerful.

feeding ur own crm sources into claude code + karpathy's llm wiki is such a big brain move honestly

agents can generate anything now but most of them still reset every session, so the real unlock is systems that actually retain and structure knowledge over time. been trying this with llmwiki and it feels closer to that direction

this is the pattern i've been trying to articulate. the wiki isn't just storage, it's a reasoning scaffold. when the context is structured right the agent becomes capable of things it couldn't do from scratch each time.

This is what agentic AI was always meant to look like, not just answering questions, but actively building and improving your systems in real time.

I've added it to my #piagent running gemma4 locally 100% 🦾

We've actually made a similar format in terminal based of Karpathy's idea of LLM Knowledge Bases. Give it a spin and let me know what you think:

The manual cycle of 'Claude breaks rule → figure out why → add fix to CLAUDE.md' gets old after the 10th session. The real pain is it never gets shorter — just bloats until you forget what half the rules mean. Does your LLM wiki auto-consolidate, or do you still prune by hand?

Feeding it your own codebase is where it gets actually useful, not just toy demos.

31K views on a dev workflow tweet signals hungry builders more than casual interest. When technical content outperforms entertainment, it's market demand not algorithm luck. Infrastructure play.

What are Shadcn packages? @grok


