Synthesized notes from AIMM mastermind sessions. The recap is the artifact; the conversations are the practice.
77 recaps
Recap · AIMM
Months of scattered pieces—context management, token use, model optimization, skill sharing—snap together this week into one unified architecture for carrying intelligence through the environment.
Recap · AIMM Spring 2026
Months of scattered pieces — context management, token cost, model routing, skill sharing — snapped together in one session. Lou walked through the architecture live.
Stop fixing the input. The model is not deterministic. Build an external evaluation agent with a gold-standard rubric that any skill can call before publishing.
The model is non-deterministic. Even great prompts produce slop on some runs. Lou's reframe: stop optimizing the input and build a quality gate that scores the output instead.
Stop pasting documents into chat. Your folder on disk is the workspace. Claude reads files on demand. The conversation is for thinking. The folder is for memory.
The conversation is disposable. The folder is the memory. Lou walked through the context architecture that eliminates hallucination and makes AI genuinely useful for high-stakes documents.
The 50% rule: Claude's reliability degrades at half-full context. The /handoff skill is the antidote. Constraints are the answer to improvement spirals.
A model-selection discipline crystallized: match the model to cognitive load, not importance. Start in chat to validate, then promote to batch. Rewind instead of correcting.
Retrieval vs inference. Obsidian search gives you retrieval. Claude Code on your vault gives you reasoning. The taxonomy is the scaffolding Claude follows to think with your knowledge.
Recap · AIMM 2026
Lou's ambient intelligence scaffold gets industry validation, the group cracks what separates AI slop from authentic output, and Don's 16-person cohort onboarding produces 'how insightful!' emails the next morning.
What stays yours when models train on transcripts? The cohort wrestled with where the line actually moves quarter to quarter, and what work is worth doing inside that uncertainty.
Build a system that decides like you, not just sounds like you. Mine your conversations for decision instances. DSPy can optimize prompts from golden examples.
Lou walked the cohort through the system he had been quietly building, a personal knowledge vault that doesn't just store what he learns but produces articles that sound like he wrote them, complete with the personal asides he could never get AI to insert before.
From research conversation to newsletter to NotebookLM explainer to email copy to vault capture. Every conversation is either spent compute or invested compute.
A skill is just a folder with a SKILL.md file. Everything else is optional. Progressive disclosure. Build the Brand Writing Team live with quality gates and adversarial steps.
Lou maps the 8 eras of AI adoption from 2020 to today, the group wrestles with the curse of the expert, and Opus 4.6 demonstrates it can reverse-engineer unconscious judgment from natural conversation.
Guest session with Michael Simmons, the writer behind 100 million Forbes views. He walked the cohort through how Claude Code became his primary thinking tool and what changed in his writing process when he stopped treating the AI as an output engine.
Schema injection goes platform-native, Lou cancels ChatGPT in favor of Claude + Google, and the /skeptic command emerges as the highest-ROI post-generation habit in the room.
AI-native browsing tools like BrowserOS are gaining adoption. Computer-use agents produce real resul...
Lou debuted Eigenthinking framework for extracting non-average ideas from AI by mapping cognitive fi...
Unveiled GEARS Wikidata semantic mapping upgrade projecting 15-25% improvement in AI answer inclusio...
Kasimir shared multi-layer prompting system combining depth drilling, recursive refinement, and firs...
Announced Gears Alpha live with full onboarding playbook. Discussed asset hierarchy for ontology bui...
Unveiled Article Hub generated from Amy Yamada's ontology - 200 articles in 12 minutes for 63 dollar...
Explored ontology as clarity machine revealing pre-awareness symptoms clients experience. Multi-mode...
Don Back launched 24-week GEO-aligned content system already driving LinkedIn engagement. Elizabeth ...
Recap · AIMM Winter 2026
First call of 2026. Don shared the content engine he built over the holidays, a central canon, five core laws, seven pillars, twenty-four weeks of mapped output. The cohort dug into how to keep it from drifting into irrelevance.
Recap · AIMM 2025
Final session of 2025 from Taiwan — Don Back unpacks the 3-layer GEO authority architecture, LinkedIn's quiet shift to LLM indexing, and why AI finds voices, not links.
Final session of 2025 from Taiwan. Don Back presented Generative Engine Optimization framework focus...
GEO app live-launched from Thailand in a single session, the full ICH-to-FAQ JSON-LD schema pipeline, GPT 5.2's signal command, and vibe coding crossing from prototype to production.
Hands-on clinic in Generative Engine Optimization with live app deployment. FAQ page with JSON-LD sc...
AI's real leverage is collapsing cycles not replacing thinkers; shadow AI in enterprise; Elizabeth's blood-work SQL architecture; and the GEO schema tool goes live.
Iteration compression thesis showing AI's value in collapsing cycles not replacing work. Enterprise ...
Recap · AIMM Fall 2025
Kasimir built a working interface in an hour with Claude Code. Lou admitted he can no longer remember how n8n's inputs and outputs work because he just vibe codes everything now. The cohort worked out where each abstraction layer actually serves you.
Pre-holiday session: Claude Code wins and reality checks, the Perplexity → Drive → Qdrant knowledge pipeline, DORA's 71% wait time reduction, and Lou's psychographic hub for AI citations.
Vibe coding progress with Kasimir's Pinecone interface. Discussed knowledge management platform gaps...
Extended hands-on session building Custom GPT against Pinecone then pivoting to Qdrant. RAG depth di...
Two hours of live RAG debugging, 20 chunking variants × 20 retrieval strategies, and the full self-hosted VPS stack for ~$15/month.
Dirk's N8N milestone after months of persistence, Claude deep research shaving five days off a grant project, and Lou's live Qdrant → ChatGPT Custom GPT → Claude MCP production stack.
Dirk milestone with self-hosted N8N in Europe. Don shaved five days off university grant using Claud...
A Claude writing team with creative license breaks the fourth wall, the identity shift from writer to thinker, and Kasimir's five-level advisory board architecture.
Lou demoed multi-role writing team as Claude skills - researcher, strategist, writer, editor, publis...
Anthropic released Skills the week before and the internet went viral. Lou walked the cohort through what they actually are, then admitted he didn't yet see why everyone was losing their minds. The room worked the question together.
Claude Skills as Anthropic's App Store play, building a self-evolving multi-role writing team in a single skill file, and when to be an early vs. late adopter.
Claude Skills released covering reusable capability packaging. Lou built complete multi-role writing...
Turning session transcripts into published content with a multi-model pipeline, the wisdom doctrine process, and tool updates across Google, Anthropic, and OpenAI.
When one API key beats three SaaS subscriptions, the CBT coach parable revisited, a live N8N multi-LLM debate workflow, and the asymmetric East-West automation story.
Analyzed paying for multiple subscriptions versus consolidated access via API. Real conversation abo...
Hands-on clinic in Generative Engine Optimization. PsyGen app automates Ideal Customer Handbook to J...
Lou deploys a working GEO app live from Thailand, the full ICH-to-FAQ JSON-LD schema pipeline, GPT 5.2's signal command for context density control, and vibe coding crossing into production.
Vibe coding wins and learning curves. Knowledge management gap solutions discussed. DORA case study ...
DORA cuts NHS patient follow-up wait times by 71%, Lou's schema.org psychographic hub gets you cited by AI engines without backlinks, and how to monetize what AI can't replicate.
Extended hands-on session building Custom GPT action against Pinecone then pivoting to Qdrant. Live ...
Two hours of live debugging connecting vector databases to ChatGPT Custom GPTs, the 400-combination RAG strategy space, and the self-hosted VPS stack for ~$15/month.
Dirk achieved N8N self-hosting milestone on European server. Don used Claude deep research to compre...
Requirements-first as the discipline AI tools can't supply, the NHS DORA voice agent cutting wait times 71%, and the Perplexity → Drive → Qdrant knowledge pipeline.
Claude Skills announcement and implementation. Lou built multi-role writing team inside single skill...
Claude Skills as the App Store play, building a multi-role writing team in a single skill, and why the 30–40% maintenance cost rule changes every AI tool adoption decision.
A practical walkthrough of self-hosting open source LLMs, from quantization tradeoffs to Docker containers to virtual private servers. The real question underneath: when does owning your stack actually matter?
GPT-5 agents work for clicks, not for nuanced knowledge work; 400 cold emails and zero replies shows where AI outreach fails; the Hormozi $150M launch decoded.
GPT-5 agents useful for manual click tasks but not knowledge work. AI cold outreach doesn't move C-s...
Lou walked through full transcript-to-thought-leadership pipeline. Wisdom Doctrine process for rever...
Turning every conversation into thought leadership with a multi-model pipeline, reverse-engineering your own unconscious expertise, and the tool updates worth acting on this week.
Analyzed cost of multiple LLM subscriptions. Real conversation about AI economic disruption from man...
AI's economic disruption from two conference perspectives, why the CBT coach parable should worry every knowledge entrepreneur, and a live N8N build where GPT, Claude, and Grok argue.
Prompting is obsolete. Process architecture replaces it with dynamic context engineering and self-ge...
Why prompting as a skill is obsolete, meta-meta prompts that interview before generating, and building an interactive Claude artifact from idea to deployed tool in under an hour.
Kasimir's self-evolving AI memory system making Claude sound progressively more like him. Don's two-...
Kasimir's local MCP memory system that learns to sound like him, the multi-LLM pipeline that's actually working, and the vibe coding discipline nobody talks about.
Kimi's human-like writing, HeyGen's uncanny avatars, the investor persona prompt flip, Don's AI-drafts-human-rewrites workflow, and a live legal RAG app for $5–10/month.
A working session that started with a new model release and ended on a deeper question: when there's a new state-of-the-art every week, what's the discipline for not chasing every shiny thing?
The hidden maintenance tail of self-hosted AI, a three-part ROI framework, why you should hire someone to teach you, and the Dia browser as a live-context machine.
Dirk's voice bot goes live and hallucinates a salary, ChatGPT recommends him unprompted, Don uncovers a hidden ideal client persona, and the honest math on local LLMs.
Live demo of GenSpark's fact-checked presentations, three-layer hallucination prevention, breaking a GPT live on screen, and Lou's $40K legal AI project.
How to handle 25,000 contacts without breaking context windows, Kasimir's 64-variant content engine, and why telling AI the goal beats giving it a word list.
Inaugural session reframing AI from task tool to whole-business architecture engine, with live demos of prompt extraction, knowledge graphs, and boardroom AI in action.
Lou opens an impromptu masterclass tracing the evolution from one-off prompts to dynamic process prompts that interview the user, plan, and then run a generalized framework. The thesis: stop writing prompts, start designing systems that write them for you.