AIMM Session — June 19, 2025: Scaling AI Workflows and Principle-First Humanization

“If you don’t know exactly what it’s supposed to do — ask the AI. Let it figure it out. We improved humanization by 80% simply by letting the AI find its own way instead of prescribing what it should think.” — Lou, on meta-prompting vs. static word lists

30-Second Summary

This week the group went deep on three practical frontiers: scaling AI workflows to massive data (25,000 LinkedIn contacts), building sophisticated content engines that score and merge multiple creative modalities, and the surprisingly sticky human question of how to actually get paid for being a smart connector in an AI-powered world. Don Back served up a sharp observation about the tech-business translation gap — and an engineering professor in the room agreed. Lou closed with a live demo showing that telling AI why beats telling it what, cutting AI detection scores by 80% in real-time. The room was loud, the ideas were better.

1. The 25,000-Contact Problem: When ChatGPT Quits on You

Context windows are not your CRM. Dirk came in with a real-world puzzle: 25,000 LinkedIn contacts in a CSV file, and a solid plan to auto-tag each one by ICP type, language, and industry. ChatGPT kept bailing after 50–100 rows — not with an error, just a polite lie that it was done.

The diagnosis? Context memory limits. The fix isn’t to keep pleading “continue.” The fix is to rethink the architecture entirely.

Lou’s recommended stack:

Break the CSV into chunks and process row-by-row using a Make.com automation loop. Let the automation handle data retrieval and writing; push the inference work to the AI via API.
Alternatively, put the data in a Google Sheet and use ChatGPT’s Google Drive connector to read 10 rows at a time with explicit instructions like “process rows 11–20.”
For massive context needs, try Google AI Studio (aistudio.google.com) — Gemini 2.5 Pro Flash currently offers 1–2 million token context windows and has a free API tier.

Kasimir’s tip: Instead of “continue,” try “continue from where you left off” — the specificity can help the model re-anchor. And look at TypingMind (typingmind.com) as a unified AI client that lets you save sessions, organize under folders, and pick up where you left off without blowing your context.

Hot Take: If your AI workflow only works for 50 rows at a time, you don’t have a workflow — you have a prototype. The real unlock is treating AI as a reasoning engine inside an automation, not as the automation itself.

2. The Content Machine: Kasimir’s Infinite Prompt Engine

The future of content isn’t a better prompt — it’s a prompt that generates prompts that generate content. Kasimir shared a live demo of a custom GPT he’s built around the “infinite prompt” framework — and it’s one of the more architecturally elegant AI content setups we’ve seen in this room.

Here’s the structure:

Feed it context: Kasimir uploads brand identity docs, psychographic profiles of his ideal clients (mapped to MBTI/personality types), and voice guidelines — all in the GPT’s knowledge base. Key assets also get pushed to memory so they persist across sessions.
Trigger with /superscript: A single keyword kicks off a multi-stage content generation process across 4 modalities: emotional narrative, research-heavy, contrarian thought leader, and archetypal/spiritual.
Run 16 prompting filters: Each modality gets run through 16 different prompting lenses. That’s 64 content variants generated from a single input.
Score and synthesize: The system scores each output across dimensions — originality (98%), resonance (99%), impact (97%) — and then synthesizes the best elements into a single final canvas.

The prompts themselves are written in a pseudo-JavaScript style — function calls with named parameters — which acts as elegant, compact shorthand when you’re working within character limits.

Lou’s observation: “You’re not just saying ‘generate 15 canvases.’ You’re telling it what each canvas should contain. That’s the design insight.”

The public GPT (non-voice-specific version) is available at: Infinite Content Brief Generator (Commercial)

Available commands: /brief · /prompt · /lite · /score

3. Getting Paid to Connect: The Matchmaker Model

Every great network is a latent business — most people just don’t charge for it. Bali brought a live dilemma to the room: she keeps meeting remarkable people doing remarkable work in the AI space (AI literacy initiatives for 25,000 people across Africa; governance frameworks for lawyers; etc.), and she’s naturally connecting them. The question: how do you formalize that into a revenue model?

Lou’s framework distills partnership negotiations down to four value levers: Sales & Marketing, Operations (product + delivery), Administration (finance, CS), and Brand Equity. Whoever brings the most of each should own the most of the deal — but if your potential partner has significant brand equity, lead with “what works for you?” and fit into their existing structures.

For connectors and super-networkers, the specific model Lou named is the JV Broker/Matchmaker — what Jay Abraham famously called “triangulation”:

You represent both sides of a deal
You vet and curate quality on each end
You take a small percentage from each party (e.g., 10% from each)
Your value is in the introduction and the alignment, not the execution

From the chat — Don Back: “You are acting as an AGENT without the structure of an agent. Have you thought about the agent model? % of deal plus costs?”

Mazie’s live aha: This conversation sparked an entirely new business concept on the spot — a concierge dating service, where clients go through a coaching intake before being placed into a curated pool.

Resources:

boardy.ai — AI-powered super-connector app Lou shared for building a curated network fast
JayAbraham.com — extensive free material; Kasimir mentioned having a full disk of Jay Abraham content

4. The Tech-Business Translation Gap (And the Opportunity Inside It)

Brilliant engineers can’t tell you what their work is worth — and that gap is a business. Don Back came back from a morning meeting with ML faculty from the University of Alberta (just returned from Singapore working with Rich Sutton’s group) with a sharp observation: even the most technically capable AI researchers are often completely disconnected from the concept of value creation.

His words: “They know how to do the thing. They don’t know what the value of the thing is.”

The result: a hall-of-mirrors academic ecosystem where everyone optimizes for papers and grants, and genuinely world-class talent can’t be placed — because they’ve never had to think from the outside in.

Alex Flueck — an engineering professor on the call — confirmed it from the inside: his own sabbatical in industry was the single biggest reorientation of his research agenda, because it forced contact with real problems, not just peer-reviewed ones.

Lou’s framing from his telecom days: The solution isn’t to make engineers better marketers — it’s to create bridge roles (“application engineers”) who combine technical fluency with communication clarity. The translator is often where the most value lives.

Don’s closing note: He’s already sold two more contracts doing this exact translation work. “Sales is the purest of professions — it starts with a creation in the mind, and nothing happens without it.”

5. Let the AI Think: The 80% Humanization Experiment

The difference between a list of banned words and a writing philosophy is about 80 percentage points. Lou walked us through a live experiment in humanizing AI-generated content — and the core insight flipped a common assumption.

The conventional approach (popularized by people like Rob Lennon): maintain a blacklist of AI-typical words and phrases (things like “delve,” “tapestry,” “in conclusion”) and instruct the model to avoid them. It works — but only as a snapshot of today’s models. By GPT-5, the list is stale.

Lou’s alternative: Don’t give the AI a list. Give it a goal.

He prompted Claude’s extended thinking mode with nothing more than: “Make this writing sound as authentically human as possible — it should be able to fool AI detection bots.” No word list. No prescriptions. Just a principle.

Results, run through an AI detection tool:

Version	AI Detection Score
Grok humanizer (word-list approach)	~10.5% AI
Claude (no word list, principle-based)	~7.6% AI
Claude + extended thinking	~5.3% AI
Final version after iteration	~2.1% AI

An 80% reduction in AI signal — not by managing a list, but by letting the model develop its own humanization strategy.

Don Back in the chat: “I was reading this week about the emerging value in pre-AI verbal content. The analogy is to pre-WWII/pre-atomic age steel — so much AI-created text has polluted written content.” — This is the business case for authentic voice, folks. The scarcity is shifting.

Community Corner

Donald used the infinite prompt to generate a complete set of psychological affirmations — work that previously took him months. The AI incorporated NLP principles, tested for phonetic quality when spoken aloud, and produced results he called “mind-blowingly good.” Time investment: about 30 minutes.

Don Back closed a sale at his lunch meeting and signed another contract at noon. He’s working with engineering grad students and postdoctoral fellows at the University of Alberta to help them translate their ML expertise into career paths that actually pay.

Mazie is now seriously exploring a concierge matchmaking service for humans. The idea emerged live in the session from the JV Broker discussion.

Elizabeth (new member) couldn’t join this week — it’s her anniversary. She’ll be with us next week. Welcome, Elizabeth!

Try This: The Principle-First Prompt Upgrade

Pick one of your existing prompt templates — something you’ve been incrementally tweaking with specific instructions. Now try this instead:

Open a fresh thread (don’t bring your old prompt).
State only your goal, not your method.
Ask the AI: “Think deeply about what would make this goal achievable and tell me everything you’d do to get there.”
Once it responds with its approach, say: “Great. Now turn those instructions into a system prompt I can reuse.”

Then compare the output to what your original prompt was producing.