How We Do AI - Two Professionals. One Mindset.
Two AI productivity professionals discuss their local folder-based personal knowledge management system, explaining how it remains tool-agnostic across AI models like Claude, Codex, and Gemini. They address listener questions about agent orchestration, token efficiency, and the ICO framework's distinction between individual versus team-related knowledge. The conversation emphasizes that the underlying organizational principles remain constant regardless of which AI brain is pointed at the folder.
Summary
The episode opens with Tom recovering from a system crash caused by running too many displays and software, which he uses as a teachable moment about having tool-independent systems. He explains that their local folder approach means recovery is straightforward because the knowledge base exists independently of any specific tool or device.
Paco shares a related 'bleeding experience' where a runaway AI process consumed 1% of his weekly Claude token limit per minute. While waiting for it to stop, he tested OpenAI's Codex for the first time and found that pointing it at his PKA folder worked identically to Claude Code — the agents, workflows, and instructions all functioned as expected. This validated their model-agnostic folder design philosophy.
The hosts then address a viewer comment questioning why named agent personas ('specialists') are useful if they're just 'switching heads' within the same context rather than running true parallel sub-agents. Tom explains that in Claude Code, the orchestrator (Larry) genuinely launches sub-agents in parallel, but even in interfaces that don't support true sub-agents, the named specialist structure still saves tokens by only loading relevant agent context rather than everything at once. They also clarify that their folder structure keeps a single source of truth for agent instructions that any model can reference, rather than duplicating model-specific skill files.
Tom demonstrates the folder architecture: a short claude.md file that references distributed agent instruction files, rather than one bloated monolithic file. The orchestrator (Larry) delegates to specialists like Nolan (HR/agent hiring), Pex (researcher), Pen (journaling), Mac (API/MCP connections), and Silas (database architect). New agents can be hired dynamically by telling Larry what's needed. The team explains why six agents ship by default and how expansion packs can be installed for more specialized needs like front-end development or design.
Paco notes that optimizing his orchestrator to run on Claude Sonnet instead of Opus dramatically improved response speed and reduced token consumption, since the orchestrator's job is delegation rather than deep reasoning. He was able to reduce his claude.md to its minimum viable size through careful analysis.
The ICO framework is explained at length in response to questions about integrating business and personal knowledge. The hosts describe four quadrants: personal knowledge management, personal project management, business knowledge management, and business project management. They argue the key distinction is not 'personal vs. work' but 'individual vs. team' — what only affects you belongs in your local PKA folder, while information shared with or relevant to a human team belongs in separate business tools. A solopreneur's content pipeline belongs in their PKA; a published article belongs in the business knowledge management area where team members and business AI agents can access it.
Tom argues strongly that solopreneurs should adopt team-style project management from day one, treating themselves as a multi-role team. He shares his own experience launching the Paperless Movement as a side hustle and immediately installing proper project management tools, which made scaling to a team seamless. The episode closes with Paco illustrating how his article creation workflow lives in his PKA (individual responsibility, multiple draft statuses, specialized agents for drafting and image creation) but only the published final article gets pushed to the business knowledge management system as a signal for the team.
Key Insights
- Paco found that switching his orchestrator from Claude Opus to Claude Sonnet dramatically improved response speed and reduced token consumption, because the orchestrator's job is delegation rather than deep reasoning — and with a properly minimized claude.md, he never ran out of context window even without compacting.
- Tom argues that even when an AI interface cannot launch true parallel sub-agents and instead 'switches heads' between named personas, the specialist folder structure still saves tokens because it loads only the relevant agent's context rather than a single monolithic instruction file containing everything.
- Tom contends that perplexity's 'computer use' product, despite offering local folder access, puts users inside a proprietary ecosystem where they lose control — the system had no understanding of their folder structure or instructions, illustrating why owning the folder layer independently of any AI brain matters.
- The hosts argue the critical distinction in the ICO framework is not 'personal vs. work' but 'individual vs. team' — information that only affects one person (even if business-related, like a content draft pipeline) belongs in a personal PKA folder, while information shared with or acted upon by a human team belongs in separate business knowledge management systems.
- Tom claims that solopreneurs who structure their business as if they already have a team — using proper project management tools and defining roles — can scale instantly when they hire because new people join a running system rather than chaos; he did this himself with the Paperless Movement and had five people integrated without business disruption.
Topics
Transcript
[0:00] Okay. Hello everyone. Welcome back to another week. Well, here we are showing the perfect use case if something crashes to come back online. Um, I'm saying this because, well, after using the MacBook Pro M1 with well, four displays, two studio displays from Apple, 5K displays, no problems at all when it comes to [0:31] performance. But at some point, especially if you somebody who is reviewing and testing a lot of software, you reach a point where just everything crashes. And this is the moment where you will be happy to have a system in place and that a system is independent from the tool or the device that you are using. And here we are. The…
Full transcript available for MurmurCast members
Sign Up to AccessMore from ICOR with Tom | AI Productivity
Claude sucks. And here is why.
This video explores why Claude and other AI models produce inconsistent results, examining how model selection (Haiku vs Sonnet vs Opus vs Claude 3.5 Fable), effort levels, prompt clarity, and system context dramatically affect output quality. The speaker demonstrates these differences through a Venn diagram creation task and advocates for using organized folder systems with documented SOPs and code-based solutions to ensure consistent AI performance.
Claude can now prompt its own SESSIONS. I built the boss.
Claude now has the ability to interact with multiple sessions, allowing a streamlined orchestration of tasks through an agent called Larry. This integration can simplify task management and enhance productivity by enabling a single main session to oversee various sub-sessions or tasks.
I switched off Claude's memory. My folder does it better.
Tom explains why he disabled Claude's native memory feature and instead built a local folder-based system (myPKM) to maintain persistent context with AI agents. He shares an audit revealing systemic issues like contradictory documentation, forgotten rules, and parallel session conflicts, then outlines how to systematically fix these problems using the ICOR productivity methodology.
I do not download invoices anymore. My folder and Claude do it.
The speaker demonstrates how he uses AI (Claude) integrated with his local folder system to automate invoice collection for his accountant, downloading 36 invoices in 5 minutes with minimal manual intervention. He explains his AI team methodology, work streams, and how persistent local knowledge bases enable AI to handle recurring tasks independently without constant re-explanation.
Claude finished the work. Then I lost it in the chat.
A creator demonstrates a comprehensive local-folder-based system for managing AI-generated deliverables asynchronously, featuring an interactive browser interface that tracks decisions, changes, and conversations without relying on external databases or web applications. The system enables collaborative editing, visual annotations, and automatic context preservation across multiple parallel AI projects.