One Life. One Folder.
The presenter demonstrates a personal knowledge management system built entirely on a local folder structure using plain text files and Claude AI via terminal, replacing five human team members with AI agents. The system features a multi-agent team orchestrated by 'Larry' who delegates tasks to specialized agents like Pax (research), Charter (design), and Sage (writing). The approach emphasizes persistent memory, brand consistency, and tool-agnostic design over cloud-based AI chat interfaces.
Summary
The presenter walks through his 'One Life. One Folder.' system, a local plain-text folder containing 24,000 files (53GB) that serves as his CRM, project manager, journal, research database, and business backbone. The folder structure, downloadable as a free scaffold from myICOR, replicates his personal setup and that of co-founder Paco Cantero, proving it scales across multiple businesses.
The PKM folder contains subfolders for CRM (people and organizations), documents, images, and a journal backfilled to 2017 from tools like Day One, Heptabase, Tana, Notion, and Obsidian. The presenter explains that the shift to this system was enabled by Claude's computer use capabilities (originally called 'openClaw'), which allowed AI to access local files and work autonomously.
Rather than using Claude Desktop chat or Claude Cowork (which limits agents to sequential 'hat-switching'), the presenter uses Claude CLI via terminal inside VS Code, running multiple parallel sessions. He argues that Claude chat and ChatGPT are ineffective because they lack persistent context, whereas his folder provides permanent, structured memory.
The multi-agent team is orchestrated by Larry, the single point of contact, who consults an agent index file to route tasks to specialists: Pax (research via Perplexity and Brave Search APIs), Charter (infographic/image generation via headless Chrome), Sage (content writing in the presenter's voice), Vera (QA specialist), Iris (brand design), Felix (front-end development), Nolan (HR - creates new agents when needed), Mack (API/automation connections), Vita (health data interpretation), Buzz (social media scheduling via Metricool), and Scout (trend research). Each agent has its own agent.md file with specialized instructions, making the system LLM-agnostic.
The presenter demonstrates a live comparison between his PKM folder and an empty folder, asking both to create a branded social media infographic with AI productivity news. The empty Claude produces a generic, off-brand result quickly, while his PKM-connected Claude takes longer but produces a properly branded infographic (dark mode, brass accents), a LinkedIn post written in his verified voice, avatar-fit scoring, hashtags, and a privacy gate check — all aligned to documented SOPs stored in the team knowledge folder.
The system connects to external tools via API keys stored in a local env file, with security agent Vex controlling which agents access which keys via hooks. External connections include Supabase (database backend for myICOR), Perplexity, Brave Search, Metricool, Notion, and ClickUp. The presenter also demonstrates remote control of running terminal sessions via Claude mobile app.
The presenter situates this system within the broader ICOR methodology, which he and Paco have taught to thousands of businesses for over four years. He argues that AI fails inside companies for the same reason automation initiatives have always failed — organizations apply tools without first understanding their workflows and work streams. His vision is for individual experts to build personal knowledge assistants that augment their expertise rather than replacing human judgment, using AI to remove friction in information retrieval and connection rather than wholesale replacing human teams.
Key Insights
- The presenter argues that Claude Desktop chat and ChatGPT are fundamentally useless because each session starts from scratch with no real persistent context, and that the only meaningful shift happens when you build persistent memory for AI inside a structured local folder — giving you full control and LLM-agnosticism.
- The presenter claims Claude Cowork is inferior to his terminal-based setup because Cowork forces a single agent to sequentially 'switch hats' rather than truly spawning parallel sub-agents, whereas his folder rules enforce Larry as orchestrator dispatching genuine parallel specialist agents — mirroring Anthropic's own published dynamic workflow architecture.
- The presenter demonstrates that his system can autonomously hire new AI agents on demand: when Larry cannot find a capable team member for a task, he routes to Nolan (HR), who tasks Pax to research the best human skill profile for that role, then automatically generates a new agent.md file with an avatar image and registers it in the agent index.
- The presenter contends that token efficiency is achieved not through model compression but through architectural separation — SOPs and guidelines are stored as single-source-of-truth documents that agents load selectively per task rather than holding all knowledge in a monolithic prompt, preventing duplication and keeping context loads minimal.
- The presenter argues that AI adoption failures inside companies repeat the same pattern as failed automation and business process optimization initiatives — organizations apply AI on top of broken workflows without first mapping what individuals actually do daily, and that the correct entry point is always personal workflow clarity before any AI layer is introduced.
Topics
Full transcript available for MurmurCast members
Sign Up to Access