How We Do AI - Two Professionals. One Mindset.
Two professionals discuss how they leverage AI-powered Personal Knowledge Assistance (PKA) systems to dramatically improve business workflows, using a real-world example of creating a 60-slide convention presentation. They demonstrate how structured AI agent systems with local databases outperform generic AI tools like Gamma or Claude Design by maintaining full personal and business context. The session also covers their multi-agent orchestration architecture, model selection strategies, and the compounding value of consistent knowledge capture.
Summary
The episode opens with Paco sharing his experience using an AI-powered Personal Knowledge Assistance (PKA) system to prepare a 60-slide presentation for his marketing agency's annual convention of 70+ people. He contrasts this year's output — generated using his PKA combined with Claude Design for polish — against last year's effort using Gamma, calling the difference 'literally insane.' The PKA system, built on Claude, pulls from months of structured journal entries, project notes, and captured ideas to generate a coherent first draft complete with brand-consistent design, structure, and key messaging. The process took roughly six hours, which Paco acknowledges is still too long but attributes largely to the limitations of Claude Design's token consumption and speed rather than the PKA itself.
The co-host shares his own parallel experience, noting that his IPA (Intelligent Process Automation) system — which manages his iCore website — produced even better design output than Claude Design because it already contained full business context, eliminating the need to manually re-feed design systems. He references a complete website and membership platform overhaul accomplished in under 24 hours using his multi-agent setup running on Claude Opus with a 1-million token context window, at a cost exceeding $1,000 in API credits, which he argues was clearly worth the return on investment.
Paco then demonstrates his Slack-based interface to the PKA, showing how he can query his local database through a Slack channel called 'Mindset,' receiving instant threaded responses from his AI system. He explains the full capture workflow: information sent via Slack on mobile is automatically triaged, enriched, tagged, and stored as a structured journal entry — including images — without any manual effort. The system then connects disparate pieces of knowledge, such as linking Henry Ford concepts with IKEA examples, and surfaces these connections when building presentations or articles.
The hosts distinguish between their approach and simpler AI tool usage, emphasizing that the PKA evolves from Personal Knowledge Management (manual, friction-heavy) to Personal Knowledge Assistance (automated, AI-driven connections). They argue this is the critical shift that compounds over time, with the AI learning the user's thinking style, communication preferences, and business context from ongoing journaling.
The co-host then showcases his AI agent library — a set of named, role-specific agents (Larry the orchestrator, Iris the design guardian, Felix the front-end developer, Pixel for thumbnail creation, a researcher using Perplexity API, etc.) — each specialized and model-assigned based on task requirements. He explains that Larry ran a 15-hour uninterrupted codebase overhaul without ever needing to compact the conversation context, thanks to the 1-million token window and a working folder system that preserved context between agent handoffs. He also discusses using model-agnostic, tool-agnostic local folder structures as the foundation, pointing any AI model (Claude, Gemini, Codex) at the same system.
In closing, both hosts address audience questions about model selection per agent, noting they settled on Opus for the orchestrator for reliability despite cost, while acknowledging Sonnet could save costs. They also confirm using external APIs — Perplexity for web research, Gemini Nano for image generation, AssemblyAI, and MX — integrated through individual agents to keep the system efficient and avoid bloating the main context window.
Key Insights
- Paco argues that the most significant improvement in his AI-generated presentation was not the design quality but the structural clarity — the PKA organized his thoughts and key messages in a way that reflected how he naturally thinks, something he had never achieved manually in previous years.
- The co-host claims that Claude Design is inferior to a locally-contextualized agent system for design work because Claude Design requires the user to manually re-feed business context every time, whereas an IPA system already knows the full design language and brand guidelines, making it effectively incomparable.
- Paco demonstrates that his PKA system, accessed via Slack, automatically triages, enriches, tags, and links captured content — including images sent from mobile — into structured journal entries entirely on autopilot, with the AI deciding entry type, project assignment, and connections without user instruction.
- The co-host reports that his 15-hour uninterrupted codebase overhaul was completed without once compacting the conversation context, because the orchestrator agent Larry has a 1-million token context window and delegated work to sub-agents who stored outputs in a working folder, allowing Larry to maintain full project oversight from start to finish.
- The co-host explains that using Perplexity API specifically for web research — rather than letting Claude do it — consistently produces better results because Perplexity has broader web access including Reddit, while Claude frequently gets blocked from reading specific pages.
Topics
Full transcript available for MurmurCast members
Sign Up to Access