AIMM Session — October 30, 2025: The Writing Team That Sounds Like You and the Advisory Board OS

“I’m sitting here trying to write an essay about thinking, and I can’t tell where my thoughts end and the models begin. That’s not a metaphor. That’s not the hook. That’s just what’s happening right now.” — Lou

30-Second Summary

Two core questions: What does it mean to be a creator when AI can generate publishable work in minutes? And how do you build AI systems that make you smarter, not just faster? Lou walked through a full content pipeline — from technical research to polished article, audiobook, and slide deck — built on a team of Claude skills. Kasimir Hedstrom pulled back the curtain on his virtual advisory board framework: a five-level architecture that evolved from a single prompt into something closer to an operating system for strategic thinking. Throughline: the people who will thrive aren’t the ones who automate fastest — they’re the ones who think hardest before they automate at all.

Topic 1: The Writing Team That Sounds Like You

Lou demoed AI-generated content with genuine voice. Started with a Claude research paper on transformer circuits — LLMs can now introspect on injected thoughts in their prompts, detecting planted concepts before generating output. Not consciousness, but functional metacognition.

Lou used that research as raw material for a writing team built as Claude skills — modular roles firing in sequence:

Researcher — finds the angle, scores for novelty and credibility
Strategist — adds narrative tension
Writer — drafts the article
Editor — checks rhythm, clarity, places where prose goes mechanical
Publisher — outputs headline variants, target outlets (Stratechery, Hacker News, HBR), SEO tags

First draft: solid. Second draft (rewritten in a blend of Gary Provost, Seth Godin, and Lou’s own voice): better. Shorter paragraphs. One-word sentences punching through long ones.

Third version stopped the room. Lou gave the team creative license to stray from the outline. The system came back with something meta: an essay that broke the fourth wall, admitted uncertainty mid-sentence. “By the way, I almost asked AI to help me with the next section. The impulse was automatic. Instead, I’m going to sit with the discomfort and try to find it myself.”

The editor scored it 3.85/5 — below average by traditional metrics — then immediately added: “Those concerns might be the point. This feels honest in a way I haven’t felt in months. Keeping it.”

The system talked itself into keeping the “worse” draft because it was more true.

Topic 2: The Identity Shift You Can’t Skip

If AI wrote the words, is it really yours?

Lou’s answer — a reframe of professional identity. Tony Robbins didn’t write all his own books. Ted talks get coached. Articles get edited into different documents.

“I am a thinker who writes” — not “I am a writer.”

That reframe changes everything. You stop treating AI like a shortcut and start treating it like a publishing team. You front-load the thinking — the brainstorm, the creative brief, the point of view — and hand off the execution. The value you provide isn’t keystrokes. It’s the 15-minute conversation that defines what the piece is actually about.

Topic 3: Process Before Automation

Don Back dropped the line of the day: “Work out the process behaviour, practice the behaviour manually, then and only then move it into an automation. I’ve had to unwind automated messes that did not build off of clear processes.”

Lou’s practical sequence:

Run it manually until the output consistently delights you
Build it as Claude skills (so it learns across sessions)
Then wire it into N8N or another automation layer
Use the freed time to think about the next process worth automating

Automate a broken process and you don’t save time. You produce errors faster.

Topic 4: Kasimir’s Advisory Board — From Prompt to Operating System

Five levels of prompt sophistication:

Tool — A single prompt: “Create an advisory board for my business.” Works once. Doesn’t scale.
Product — A standardized blueprint in a system prompt. “Board in a box.” Reusable.
Factory — A meta-prompt that generates board prompts on demand, for any domain.
Adaptive Factory — Three modes: interactive co-creation for founders needing guidance; one-shot generation for experienced users; red team design mode for anti-fragility testing.
Operating System — Adds an owner’s manual, three-step human protocol, integrity checks to prevent “persona collapse” (when an AI board member forgets its role and goes generic).

Core insight: a board where everyone agrees has zero strategic value. Kasimir intentionally designs for productive conflict. Different experts hold opposing views. The system routes questions to 3–4 of 16 potential board members based on relevance, then synthesizes the disagreements into a higher-order strategy neither expert would produce alone.

Five-vector framework for selecting board members:

Did they create something that redefined their field?
Are their principles still valid decades later?
Did they achieve results with their own methods?
Do other experts consult them?
Did they bridge disparate fields into something new?

Topic 5: What Happens When AI Catches Itself Thinking

Researchers could inject specific concepts directly into an LLM’s neural activations — mid-process, before token generation — and measure whether the model was aware of the injection. It was. It could detect planted concepts, evaluate them, and choose to use or ignore them in output. That’s something that looks structurally like metacognition.

Implication: when you build elaborate prompt systems, you’re not just giving instructions. You’re creating the conditions for a reflective process.

Hot Takes

Lou’s writing team isn’t static. After every session, the system saves what it learned. Quality compounds. This is what separates a skill from a prompt.
Kasimir’s 3–5 Iteration Rule: after 3–5 refinement loops, ROI on prompt improvement goes negative. Changing commas doesn’t justify rewrites.
On Persona Collapse: the biggest failure mode in multi-agent systems isn’t hallucination — it’s drift. Board members forget their roles, become vague, start agreeing. Kasimir builds integrity checks directly in: commands that reset a persona to its original definition when it starts collapsing.

Try This Before the Next Session

Pick one piece of content needed in the next two weeks. Don’t automate it. Work through it manually with these five roles in sequence — even as five separate Claude conversations:

Research: what’s novel and credible? Score it.
Strategy: what’s the narrative tension? What does the reader need to feel, then resolve?
Write: draft it. Don’t edit while writing.
Edit: read aloud. Mark every sentence that sounds like a press release.
Publish: write the headline, sub-head, and name one specific outlet or audience segment.

When done manually three times, you’ll know exactly what the skill should do. Then build the skill.