Closed my first real retainer last month and I start in three weeks!! Putting the whole plan here because I'd rather find out now what I got wrong than in month three The situation. Family holding in LatAm, retail plus financial services, several companies under it. Monthly retainer, four hours a day on site, three months to start. What they want is the stuff that lives in three or four people's heads: why decisions get made the way they do, what got tried before, what broke. Written down, and answerable in plain language. The long game is a private system per founding partner, for succession. I'm one guy. Nothing signed yet... The plan for the first 3 months: Month 1 is entirely non-technical. Map where the knowledge actually lives, which folders, which spreadsheets, whose head. Then measure a baseline before touching anything: how long it takes someone to find an answer today, how often the person who knows gets interrupted. Without that number, nothing I build later can be compared to anything Month 2, a small non-sensitive corpus goes in and a second real user starts using it. Then a hand-written eval set built with their own person. Twenty to thirty questions, expected answer, expected source document, plus negative controls where the right answer is "no record of that" Month 3 is the part I care about. Week 11 I don't open the system at all, and we count how many queries happened without me. That's the acceptance criterion, written into the contract: the designated users operate it without the consultant present. Hard to fake on purpose Three calls I made that I'm not certain about: Cloud with a zero-retention agreement instead of on-prem, even though they asked for maximum privacy. My reasoning: on-prem means weeks of importing hardware before anything exists, the models that fit in an office server are worse, and buying only makes sense once you've measured real usage. So the architecture keeps the corpus and the index on their side and treats the model as a swappable endpoint. Reversible in that direction, not the other one

Over the past weeks I completed an AI Governance Professional course that also covered agentic AI systems, and I turned the learnings into working as a governance professional for the Swiss and EU context (revDSG, GDPR, EU AI Act). The thing that carried the pivot is an ICM workspace. The factory (Layer 3) is a corpus with two shelves. One holds the legal texts: Swiss data protection law, the EU AI Act, GDPR, regulator soft law. The other holds the Responsible AI stack: NIST AI RMF, ISO/IEC 42001, the OECD AI Principles and Due Diligence Guidance, the Council of Europe HUDERIA methodology. Every folder carries a MANIFEST declaring source, date and license. On top of the corpus sits the layer I care most about, because it moves from citation to action. Cross-walks map the frameworks against each other in delta tables, for example ISO 42001 against the AI Act, or the OECD due diligence steps against the Swiss and EU legal stack. Mitigation playbooks take recurring constellations (cloud LLM with personal data, professional secrecy plus external AI, high-risk systems needing human oversight, the first 24 hours after a suspected breach) and translate the obligation into contractual, technical and organisational measures, with anti-patterns and escalation triggers. Templates chain into a full workflow: intake, triage, a lifecycle risk register along the NIST phases with a named owner for every risk, threat model and red team plan, then DSFA and FRIA where required. Everything ships in German and English, and every measure links back to the corpus anchor it operationalises. It bends the ICM shape in two places. Router instead of pipeline. A compliance practice has no single sequential flow, so my Layer 1 is a routing table: task type, then workflow contract, then which corpus files to load. Same layered context loading, different topology. Provenance as a hard rule. Every substantive claim the agent makes must cite a stable block anchor into the corpus (#^art21-abs3 resolves to Art. 21(3) Swiss DPA). Anything outside the corpus gets flagged as a knowledge gap to check externally. A /sync drift check compares the MANIFESTs against what actually sits on disk. In a field where one hallucinated citation can end a client relationship, this rule builds more trust than any disclaimer.

Writing this with love. Before AI, I trained as an anthropologist, studiyng cultures, societies and cults from the Horn of Africa to India and central London. So the other night, mid check-in, I opened a file I have used professionally before: the cult checklist I used for another research on a cryptocurrency called HEX. And then I ran us down it honestly. ✅ An origin myth. Ours is genuinely good: a Marine who wired cryptographic systems into fighter jets, went to Edinburgh, and came back down the mountain holding a folder structure. ✅ A charismatic founder. Jake, with love. ✅ A sacred text. We call it "the ICM paper," as though there is only one. ✅ A cosmology. There is a force of chaos in our universe and its name is context collapse. Against it we hold a single ward, a file called CLAUDE.md, and we are all quietly certain it is working. ✅ An initiation you cannot skip. The classroom is built like the method itself: numbered, sequential, load-bearing. There is even an original sin, and it is skipping the Foundation. The confused always turn out to have skipped the Foundation. ✅ A rite of testimony. Half the intro posts open with some version of "it finally explained the why," which is one candle short of a conversion story. ✅ A rival tribe. The Framework People. The LLM wikings. May they orchestrate in peace. By the anthropology, we tick every box. We are a cult. Here is what the checklist hides, though. So is everything you have ever loved. Founder, scripture, cosmology, initiation, testimony, a rival tribe: that is not the anatomy of a cult, it is the anatomy of belonging. "Cult" is mostly just the word for someone else's belonging, seen from outside the fence. I know the difference because I have stood inside a real one. Years ago I studied the world around HEX, a crypto currency, whose members call themselves Hexicans, and whose devotion ran on a mechanic that locked your actual money away for years in exchange for a promise. Their founder was also a bit out of this world (to say the least). That is where the joke turns into a compliment, because the thing that makes a community dangerous was never the jargon or the founder. It is what the commitment costs you, and who collects.

I have been thinking about this recently as we set a carbon emission baseline at the NPO I work for. We know the environmental impact of data centers, and also the positive examples. I have built partnerships with data centers in Switzerland that sit above 1000m sea level (above 3200ft), where it's cold enough to skip most of the cooling, run on renewable energy, and use the generated heat to supply the local cheese factory 🧀 We also have numbers for the water and energy behind a single chatbot conversation, although honestly those vary wildly depending on who's counting and what they count. But does - and if yes how - agentic AI break that?. One request isn't one conversation anymore. It's tool calls, retries, sub-agents, all invisible to the person who asked. So the per-message figures we've been quoting don't really add up to anything. Does anyone have experiences, concrete numbers or case studies on this? I'm aware I could have asked AI, but wanted to start a conversation here first 😅

It’s live. Starting today, everyone in Clief Notes has access to the new Clief Notes AI. The easiest way to use it? Don’t overthink it. Ask it the question you would normally ask me. “Where should I start?” “What should I focus on next?” “Where did you talk about [topic]?” It’ll use what’s already inside Clief Notes to help answer you and point you toward the right lesson or resource when there’s something worth going deeper on. The goal isn’t to give you another AI tool to play with. It’s to make everything already inside this community easier to actually use. If you want Access to it Comment "READY" and we'll send you access to it!