AIMM Session — September 4, 2025: Vibe Coding, Voice Agents, and the Knowledge Pipeline

Session Overview

This session covered Claude Code and vibe coding in practice, a replicable knowledge pipeline architecture, an NHS healthcare voice agent with measurable patient outcomes, the psychographic hub for GEO authority, and the concept of the “sourdough excavator” — business value hiding where AI will never fully automate.

Vibe Coding in Practice: Requirements First Is the Discipline

Claude Code and vibe coding tools are powerful — but the skill they can’t supply is knowing what to ask for. Lou’s framing: the real workflow is:

Talk to the AI until your requirements are right
Ask it to generate a requirements document
Feed that document to the coder
Implement one thing at a time

Skipping the requirements step leads to rabbit holes that eat your day. The lesson: knowing what questions to ask is the craft. The code is the easy part.

Lou also mentioned an open-source orchestrator that lets you run 5–6 Claude Code sessions concurrently on the same codebase without them overwriting each other — a parallel agent setup worth testing.

Stack context: the abstraction hierarchy runs raw code → CLI tools like Claude Code → N8N → Make.com → packaged apps. N8N now has a built-in AI builder, but Make.com remains the easiest entry point for no-code builders who still want to understand data flow.

The Knowledge Pipeline: Perplexity → Drive → Qdrant

Don named the problem everyone feels but rarely articulates: ChatGPT has built up a rich history of everything he’s worked on — but he can’t find a damn thing in it.

Elizabeth’s gem: Perplexity’s artifact view lets you save outputs directly to Google Drive (or download as .md) without copy-pasting. She used Perplexity’s Deep Research mode to find updated versions of the Infinite Prompt Generator, compare old vs. new, synthesize a hybrid — then turned it into a new Claude skill.

Lou’s proposed pipeline: Perplexity artifact → Google Drive → N8N trigger (3 simple modules) → Qdrant vector store. With Google’s built-in file search, you don’t even need to manage embeddings manually.

Kasimir added: Claude’s memory feature (opt-in under Settings → Personalization) now lets you reference previous chats directly. No more handover documents.

The unsolved gap: None of the current platforms are designed as repositories. Everything valuable ends up either lost in chat history or manually copied out. The community consensus: this is a product gap waiting to be filled.

Voice Agents in Healthcare: The DORA Case Study

Bally shared a visit to a UK university project that deployed an AI voice agent called DORA for post-cataract surgery follow-up at an NHS trust.

Results:

35 weeks → 10 weeks patient follow-up cycle time
92% patient satisfaction — from an elderly population who knew they were talking to an AI
Built by a startup using WAPI, with applied linguistics researchers from Oxford and Newcastle shaping the conversational design

The key engineering wasn’t the AI — it was the linguistics. Researchers tracked interruptions, latency, filler words, and cultural communication patterns.

Don’s context: He spent years as VP of Health Innovation for an Alberta agency. Voice agents win because: staff resist outbound calls (repetitive, draining, low-status work) / bots never fatigue / patients can’t leave a message — bots eliminate the “call back later” dead end / multilingual capability serves immigrant populations without misinterpretation risk.

Don, in chat: “The Alberta government is experimenting with using ChatGPT to draft legislation. What could go wrong?”

The Psychographic Hub: GEO Without Backlinks

Lou unveiled what he called “the single most powerful authority play we’ve ever had in the mastermind” — a collaboration with Ken Droz (of the Ideal Client Handbook platform).

The insight: Old SEO is dead. The new game is showing up in generative engines — ChatGPT, Perplexity, Grok, Gemini. The mechanism: the knowledge graph, specifically schema.org structured data embedded in a single HTML page.

What the tool does: Takes your Ideal Client Handbook (ICH) as input → Extracts every fear, belief, value, desire → Generates ~400 long-tail questions your exact client is typing at 3 AM → Answers each one in your voice → Embeds entity signals linking to Wikidata → Wraps it all in schema.org markup.

The result: you paste one code block into a single HTML page on your site. No visible content required. No backlinks needed.

The Sourdough Excavator: Monetizing What AI Can’t Replicate

Starting from a reading on flow states and mind-wandering, Lou prompted Grok to make cross-domain connections, surfacing divergent ideas. The output became a 2,000-word article.

The core idea: “Monetize Your Bias.” The analogy: sourdough bread. What makes sourdough more valuable than commercial bread is the introduction of a contaminant — something that doesn’t belong in the ecosystem — that creates a new and valuable product. Your unusual element, your weird perspective, your hard-won scar tissue — that’s your sourdough starter.

The Flash Boys parallel (via Don): Mike Novogratz made 8 figures annually arbitraging the spread between buy and sell. Then algorithms got faster, and the spread collapsed to nothing. The information didn’t disappear — the gap did. We’re standing in the same room.

Lou’s reframe: AI converges expertise to the average. The value is at the edges. What AI can’t yet do: tell a client “that won’t work — I’ve tried it” / apply pattern recognition from 20 years of watching launches fail / know what not to do.

Try This Before Next Session

Run the Sourdough Test on yourself. Open a new Claude or ChatGPT conversation and paste this prompt:

“I want you to act as a Socratic interviewer helping me find my differentiated positioning as a knowledge entrepreneur. Ask me one question at a time about how I see problems differently than my peers, what mistakes I’ve watched others make that I’ve avoided, and what I know that isn’t obvious even to other experts in my field. After 5–6 exchanges, reflect back what seems genuinely unusual or irreplaceable about my perspective. Don’t summarize — push me to be more specific.”

Then bring your answers to the next session.