Claude Code + Karpathy's System = $10,000 Skills
Jack Roberts explains how to build 'super skills' in Claude Code using Andrej Karpathy's four guiding principles as a foundation. Unlike basic markdown skill files, super skills feature persistent memory, self-improvement loops, and real data connections. The video walks through a four-pillar methodology: proper skill creation, data connectors, a memory operating system, and a refinement loop.
Summary
Jack Roberts opens by claiming that most people misuse Claude Code's skills feature — treating them as static markdown files that lack context, memory, and improvement mechanisms. He contrasts these with what he calls 'super skills,' which are built on Andrej Karpathy's mental model framework and share three core properties: indexed and recallable conversation history, access to job-specific tools, and self-scoring output that improves over time.
Roberts distinguishes between two types of skills: utility skills (simple, single-purpose automations like URL shortening via Bitly) and super skills (complex, evolving systems like an AI-powered signal dashboard that aggregates news from Anthropic, OpenAI, Y Combinator, and Google). He argues that most skills fail because they are generic, static, and amnesiac — they lack strategic context, never update, and reset completely between sessions.
The Karpathy foundation Roberts references comes from a GitHub repository (88,000+ stars) inspired by Karpathy's tweet outlining four principles for better AI model outputs: think before coding, simplicity first, surgical/orthogonal edits, and goal-driven execution with verifiable success. Roberts shows how to install this repo directly into Claude Code so these principles are embedded in every session.
He then walks through his four-pillar super skill methodology. Pillar one is proper skill creation: rather than writing skills by hand, Roberts recommends prompting Claude with a detailed intention statement and having a back-and-forth conversation to define scope, tools, output format, and data sources before the skill is generated. Pillar two is data and connectors: skills need 'eyes,' and Roberts shows how to use Claude's built-in connectors (e.g., Gmail, Figma) as well as custom MCP connectors like Firecrawl (preferred for token efficiency and AI-optimized scraping) and Zapier as a universal bridge to any external platform not natively supported.
Pillar three is a memory operating system, which Roberts describes as having three buckets: long-term conversation archives (append-only wrap-up sessions stored via Pinecone vector search), immutable knowledge bases (YouTube videos, books, community posts), and a mutable current-strategy profile (one markdown file Claude reads every session). He offers a free downloadable memory OS skill from his School community, which includes three commands — wrap-up, recall, and strategy awareness — and generates a visual HTML dashboard of memory state, active strategy, and session history. Roberts argues Pinecone is preferable to Obsidian for scalability, since Obsidian's token burn increases with file size.
Pillar four is the refinement loop: after using a skill, the user grades its output, provides feedback, and Claude updates the skill's own code accordingly. This means each subsequent run is better than the last, and all long-term value accumulates through these iteration cycles. Roberts closes by mentioning a full Claude Code course covering everything from setup to monetization, available in his paid community.
Key Insights
- Roberts argues that most Claude Code skills fail for three structural reasons: they are generic markdown files with no strategic context, they are static and never update as a business evolves, and they have no memory — resetting completely each session as if the previous conversation never happened.
- Roberts claims the Karpathy GitHub repo (88,000+ stars) distills four principles — think before coding, simplicity first, surgical edits only, and goal-driven execution — and that embedding these principles directly into skill files, not just coding sessions, fundamentally changes output quality.
- Roberts states that skills need 'eyes' and recommends Firecrawl over native web browsing specifically because websites are optimized for humans rather than AI, and Firecrawl reduces token costs while delivering cleaner, more AI-readable data.
- Roberts describes a three-bucket memory architecture where long-term conversation archives and immutable knowledge bases live in Pinecone vector search, while a single mutable markdown file holding current strategy is read by Claude every session — giving skills both historical depth and present-tense awareness.
- Roberts argues that all the real value in a super skill accumulates through the refinement loop — where the user grades output, gives feedback, and Claude rewrites its own skill file — meaning the skill literally gets better with every use rather than remaining static forever.
Topics
Full transcript available for MurmurCast members
Sign Up to Access