The knowledge atlas LEN P. VAN DER HOF

Good ideas are connected.

Explore the connections between AI, search, ventures and the strategic self. Find an idea. Follow a thread. See what connects.

Explore the atlas Latest articles
Articles · Concepts · Books · Projects Follow by RSS

The knowledge atlas

Grows with every published article
52 articles 31 concepts 751 connections

137 results/751 connections

Drag to explore · Scroll to zoom · Select a node
Index

Articles · 6 min read

AEO vs GEO vs SEO: three jobs, not three synonyms

SEO, GEO, and AEO are three jobs. SEO is ranking in ten blue links: crawl, index, rank. GEO, generative engine optimization, is the system around an answer engine: retrieve, interpret, extract, verify. AEO, answer engine optimization, is the craft of one URL so that engine can lift a clean capsule. GEO without AEO is a crawl of pages nobody can quote. AEO without GEO is a beautiful capsule the crawler never fetched. They run in parallel, not as a ladder, and they produce three separate numbers: rankings in Search Console, crawler fetches in your own access logs, and named citations in a prompt set you logged yourself. Authorized Economic Operator and geography are collisions, not this page.

Articles · 4 min read

An agentic workflow is not an automation

An automation runs a fixed path. Ticket in, fields filled, Slack ping. The diagram does not invent a new diagram. An agentic workflow holds a goal, uses tools, and may change the next step when a tool returns something messy. That is useful only if someone wrote a charter, named the human checkpoints, and set an authority ceiling. Without those, you did not design a workflow. You put a junior hire with root on a looping path.

Articles · 6 min read

AI agent or chatbot: three tests before you mix the words

An AI agent is a system that takes a goal, uses tools, and acts across more than one step without a human prompting each one. A chatbot answers a turn. You type, it replies, and the work dies when the tab closes unless you saved it. An automation runs a fixed script. Mix agent and chatbot and the chatbot inherits an API key. Three tests: does it hold a goal, use tools, and continue across steps? If the answer is no, no, no, you have a chatbot. That is a respectable product. It is not an agent. An agent is only accountable with a charter: inputs, outputs, an authority ceiling, and a named reviewer who can fail the work.

Articles · 3 min read

AI Agents for Startup Strategy: Build the Strategic Intelligence Layer

AI Agents for Startup Strategy teaches founders and operators how to deploy agentic AI as a strategic intelligence layer. Using the SENSE framework, five named agent systems, versioned claims, kill criteria, and a 90-day plan, it turns tactical AI use into governed strategic sensing for startups of roughly five to two hundred people.

Articles · 16 min read

How Answer Engines Choose Sources

Search and answer engines can cite a page only when they can access it, understand its subject, extract a relevant answer and trust the supporting evidence. The strongest strategy is therefore excellent technical SEO plus original, well-structured, source-backed content.

Articles · 4 min read

If the answer is in paragraph nine, you did not write a page

Answer-first writing means the first thing a tired reader can use is the conclusion, with the conditions attached, before the origin story, the market size, or the suspense. It is not a robot voice. Google says you do not need to rewrite for generative AI, you do not need to chop a page into chunks, and there is no ideal length. AEO is this craft on one URL. GEO is the system around it. Write the sentence a skeptic can check. Then earn the right to be long.

Articles · 5 min read

What is BooksOS? Seven gates, not a writing CMS

BooksOS is a gated production system for serious nonfiction written with AI agents. The seven stages are Brief, Bible, Map, Draft, Edit, Package, Launch. Each stage owns one decision and ends at a gate. Agents do specialist work inside the gates. Judgment, voice, and accountability stay human. A fluent draft that cannot survive a fact-check is not a book. It is inventory you will throw away. The operating system is the gates, not the model that drafted the chapter. It is not a commercial CMS and not a generic write-faster-with-AI course. The book that owns the name is The Agentic Author, which is live.

Articles · 6 min read

Cialdini in the funnel: six levers, one honesty test

Robert Cialdini's Influence (1984) named six principles of persuasion: reciprocation, commitment and consistency, social proof, liking, authority, and scarcity. In Pre-Suasion (2016) and in the expanded Influence (2021) he added Unity: people are more ready to say yes to someone they treat as one of them. A marketing or sales funnel is a sequence of those asks. You benefit by knowing which lever you are pulling on each step, and by using only the ones you can keep when a human answers. A genuine sample is reciprocity. A counter that invents '12 people viewing' is a claim. The published honest-conversion rule still holds: automate the follow-up, not the claim. This page is the map of the levers. It is not a script for pressure.

Articles · 3 min read

Citation-readiness audit: eligible is not cited

A citation-readiness audit is a written fail/pass on one public URL. Four checks, in order: can OAI-SearchBot fetch it, is one checkable claim in the first HTML screen with a person or organisation attached, did you test ChatGPT Search rather than a memory chat, and did you log named / linked / used-not-cited as three different outcomes. A page can be eligible for ChatGPT Search and still never be named. Robots.txt, schema, and llms.txt do not close that gap. Google's generative reports measure Google, not ChatGPT. This page is the card; the how-to and the diagnosis are separate pages.

Articles · 4 min read

Content architecture that answer engines can lift

Content architecture for answer engines is not a schema trick. An engine can lift a page only when the claim is stated once, on one URL, attached to a named person or organization it can verify, and backed by evidence it can see. GEO is the system. AEO is the page. CITATION is the standard: discoverable, attributed, answer-shaped, evidenced, and actually used. Structured data can name the entity. It does not guarantee a citation. FAQPage is not a current rich-result tactic. If the same idea lives in six half-pages, the engine will quote someone who wrote it once.

Articles · 8 min read

Downstream effects: why the real cost of a decision lands later, and somewhere else

A downstream effect is a consequence of an operational decision that is delayed, lands on a different surface than the metric you optimized, and had no owner in the room when the decision was made. Downstream consequences is the same river. Downstream affect is the verb; the noun you want is effect. You do not need to predict every consequence. You need a way to name the next bend before commitment.

Articles · 4 min read

Evaluate the reasoning, not the fluency

Fluency is not evidence. A language model can return a clean paragraph that invented a number, cited a blog as a study, or retrieved the wrong chunk with a straight face. Evaluating the reasoning means scoring the trace: which sources were gathered, whether you can name the passage, what assumption is load-bearing, and how strong the claim is allowed to be. CALIBRATE is that match of confidence to evidence. GRAIN is the retrieval loop that keeps an answer tied to a chunk you can point at. If you cannot inspect either, you do not have evaluation. You have a vibe.

Articles · 4 min read

Write the five lines before you act

A five-line decision journal is Artifact 3 in The Psychology of Founder Resilience. Write it before you act. The live lines are: Decision (what you choose and what it forecloses), What I believed, What I feared, What happened (fill at review), Next time (one process adjustment). An entry after the result is compromised, because hindsight rewrites what you knew. The book's own evidence note is honest: a journal does not guarantee better decisions. It separates process from outcome so later-you can score the call. CALIBRATE is the practice of matching the strength of a judgment to the strength of the evidence. The Calibrated Mind is live in English and Dutch. The five lines are already usable.

Articles · 5 min read

Flow is a condition of work, not a brand mood

Flow, as Mihaly Csikszentmihalyi described it, is a state of deep involvement in an activity: challenge and skill in balance, clear goals, immediate feedback, and the work done for its own sake as well as for the outcome. He wrote Flow in 1990 after earlier work in Beyond Boredom and Anxiety (1975). In business it is not a vibe for the brand film. It is a condition you can design for in how people work, how a product teaches the next step, how a sales conversation is a craft, and how a founder day is sequenced. If the task is far too hard, you get anxiety. If it is far too easy, you get boredom. Marketing that promises 'flow' and then interrupts every twelve seconds is selling the word and blocking the state.

Articles · 4 min read

A holiday is not a refill protocol

Founder capacity is infrastructure you train and budget, not a mood that returns after a holiday. The live RESILIENT capsule names the work: sleep, recovery, decision quality, and the social cover that lets a founder tell the truth. Treat those four with the same seriousness as the cap table. RESILIENT is an authorial dashboard, proposed and applied, not a validated clinical scale. The Psychology of Founder Resilience is live. This page is the spec, not the book introduction and not a second weekly-review article.

Articles · 3 min read

Founder systems you can run when willpower is gone

A founder system you can run when willpower is gone does not ask you to try harder on a Tuesday at 23:00. The Psychology of Founder Resilience treats willpower as a folk model, not the instrument. The live replacement is a dashboard you read and a weekly review you do not skip: which RESILIENT gauge is empty, and what decision should be deferred. RESILIENT is an authorial dashboard, not a validated clinical scale. Leadership Under Load treats prolonged pressure as a design condition; reducing the load is a valid answer. The Attention Architect treats the inbox as architecture, not a sort you do with leftover grit. The first book is live. The other two are in production and not for sale. The cadence is already usable.

Articles · 3 min read

GEO meaning: generative engine optimization, not geography

GEO, generative engine optimization, is the work of becoming a source that answer engines can retrieve, interpret, extract, and verify. It is not geographic SEO. Ranking in ten blue links is a different job. Eligibility, an extractable answer, original value, and evidence that survives scrutiny are the levers. Google does not require schema, chunking, or an llms.txt file for AI Overviews. A page that cannot be quoted cleanly will not be cited, even if it ranks.

Articles · 5 min read

A go-to-market system that survives the first ten customers

A go-to-market system that survives the first ten customers treats those customers as evidence, not as trophies. Ten is the working set The Founder Sales System uses for deals closed on instinct. It is not a statistical threshold. If you have four, audit four. Write one sourced claim per customer: who they are, why they bought, what they use, what would make them leave. SENSE is how that claim enters a standing layer: Scan, Evaluate, Navigate, Simulate, Execute. PREDICT is how you keep the prior visible so the tenth customer cannot rewrite what you believed before they paid. AI Agents for Startup Strategy is live. The other two books are in production and not for sale. None of this is a vendor CAC story.

Articles · 6 min read

HITL in AI workflows: put the named person on the irreversible step

Applying human in the loop to an AI workflow means a named person sits on the irreversible step with a written fail line. Charter the agent. Write the override doctrine. Put the reviewer on send, spend, delete, and speak-as-the-company. Show that reviewer the exact payload, target, reason for action, and reversal cost on one screen, with a deny that stops the call. For a new write tool, inspect the first ten attempted writes end to end. That startup review does not retire permanent gates: sends, spend above the ceiling, deletes, and company speech still require the named person. If the output can leave while that person is offline, the loop is broken. A dashboard after the send is archaeology.

Articles · 4 min read

A conversion system that stays honest when you automate it

A conversion system that stays honest when you automate it treats the sequence as a clerk, not a closer. Automate the follow-up: the reminder, the unanswered question, the date you already named. Do not automate the claim. Fake countdown timers, manufactured scarcity, shame in the subject line, and a testimonial the customer did not write are not 'lifecycle marketing.' They are a faster way to spend trust. TRUST is the pre-send test from Marketing Without Manipulation: Transparency, Respect, Utility, Selectivity, Time. PRICE is the honesty of the number and the frame. VOICE is whose name the from-line still carries when a model wrote the email. All three books are in production. None is for sale. The send rule is already usable.

Articles · 10 min read

How to get cited by ChatGPT

There is no published ChatGPT citation ranking. OpenAI's publisher FAQ says any public site can appear in ChatGPT Search, and that summaries and snippets require you not to block OAI-SearchBot. That is eligibility. To be the source named in the answer, put one checkable claim in the first HTML screen, attach a person or organisation the machine can identify, and keep the evidence on the same URL. GPTBot is a separate training switch. ChatGPT-User is a person-triggered fetch and does not decide Search inclusion. Measure appearances, chatgpt.com referrals, and conversions as three numbers.

Articles · 6 min read

Human in the loop meaning: a named person who can fail the work

Human in the loop means a named person holds an authority ceiling and a review gate. That person charters the work, can stop a tool call, and owns the outcome. It is not a human glancing at a dashboard. If no one can fail the output against written criteria, the loop is decorative. A vendor slide that says HITL without a named reviewer is a slogan. The name, the ceiling, and the fail criteria have to live in the runbook, not in the deck.

Articles · 4 min read

Introducing Undominated.ai: the independent AI inference price index

Undominated.ai is an independent AI inference price index. It ranks models on independently measured capability first. Price breaks ties. A model that is both worse and dearer is named as dominated. Unrated is not zero. On 25 August 2026 the live board held 409 models from 50 providers; of 108 rated, priced rows, 97 were beaten on quality and undercut on price, and 11 were not. Those eleven are the value frontier. The site takes no cut of inference, no affiliate, and no paid placement. Snapshot numbers move. The live board is the source.

Articles · 5 min read

Kahneman for marketing: System 1 is not a permission slip

Daniel Kahneman's work, much of it with Amos Tversky, is a map of judgment under uncertainty. Prospect theory (Econometrica, 1979) codes outcomes as gains and losses from a reference point. The value function is generally steeper for losses than for gains. Thinking, Fast and Slow (2011) popularised two modes of thought: a fast, associative System 1 and a slower, more effortful System 2. Marketing lives in that map every time a price, a plan, or a headline sets a reference. You benefit by designing for how people actually compare, not for the rational agent on the whiteboard. You do not benefit by treating System 1 as a door you are allowed to sneak through. A frame that would collapse if the buyer had ten quiet minutes is not applied psychology. It is a rush.

Articles · 5 min read

Literary agent or AI agent: two different jobs

A literary agent is a human who represents an author's work to publishers, usually on commission, and is not this website. An AI agent is a system that takes a goal, uses tools, and acts across steps under a charter. If you searched 'literair agent non fictie' looking for someone to sell a Dutch nonfiction manuscript, this page is a collision warning, not a roster. If you searched because a slide mixed the words, three tests still decide whether the software is an agent: goal, tools, continuation. This site publishes operator nonfiction. It does not shop manuscripts.

Articles · 4 min read

LLM integration is an operating-design problem

LLM integration is not an API problem. The call is cheap. The scarce work is operating design: who owns the model and the prompt, what the tools may see and change, who can fail the output, and whose name it carries. A named reviewer and a written override doctrine are the integration. STACK versions the artifacts. MCP is the contract on one tool. Neither substitutes for those four seats being filled.

Articles · 6 min read

Marketing nudges are choice architecture, not a growth hack

A marketing nudge is choice architecture applied to a commercial decision. Richard Thaler and Cass Sunstein, in Nudge (2008), treated a nudge as a change in how options are arranged that predictably shifts behaviour without forbidding options or significantly changing economic incentives, and that stays cheap to avoid. Fruit at eye height is a nudge. Banning the pastry is not. In a funnel the default, the order of fields, the extra click, and the pre-ticked box are the architecture. You already have them even if you never named them. The work is to design them so a tired buyer can still see the choice and reverse it. A default that hides the price, invents a limit, or makes opting out a maze is not a nudge. It is a claim you would not defend on a call.

Articles · 5 min read

MCP explained for founders

MCP, the Model Context Protocol, is a contract for how a language model calls tools. The contract names what the model may see, what it may change, and how each call is logged. A founder should treat an MCP server like an API with production credentials: no anonymous tools, no undocumented write paths, no server that can spend or delete without a ceiling. Without that contract you do not have integration. You have a hole in the repository.

Articles · 4 min read

MCP meaning: Model Context Protocol, not a product

MCP means Model Context Protocol. It is a contract for how a language model calls tools: what it may see, what it may change, and how each call is logged. The official split is host, client, and server. A server exposes tools, resources, and prompts to a host: tools are model-controlled, resources application-controlled, prompts user-controlled. The official reference servers are filesystem, fetch, git, memory, time, sequential thinking, and everything. They are educational examples, not a production catalogue; the other former reference servers are archived. The MCP Registry is the catalogue. ReasonKit is first-party to this site, not an official reference. MCP is not a product you buy or the definition of an agent. A clean server with no ceiling is still a hole with nicer plumbing.

Articles · 3 min read

What an agent may never do

Override doctrine is the written list of decisions an agent may never take. The live glossary names four: spending above a ceiling, speaking as the company, changing production data, and closing a judgment the founder still owns. It is the complement of a charter. The charter says what the agent does. The doctrine says what it must refuse. Write it before the agent gets a tool that can send, spend, or delete. A named human in the loop is not a substitute. If the irreversible act already happened, the loop is decorative. CHORUS Review is the gate that enforces the list. The list is not the six-skill protocol.

Articles · 3 min read

Positioning is a decision, not a tagline

Positioning is a decision about who you will not serve and what you will not claim. A tagline is a sentence that can still fit everyone. A position is a refusal you can keep at the table, on the homepage, and in the company you would still want to run. The walk-away is part of the position: if you will take any logo and any discount to keep the conversation, the homepage is lying. The Negotiation Operator, The Meaning Operator, and The Authentic Operator are in production. None is for sale. The two lines are already usable: who you refuse, and the claim you will not make.

Articles · 3 min read

The Psychology of Founder Resilience: Capacity Is Infrastructure

The Psychology of Founder Resilience is an operator manual for high-stakes founders who need trainable psychological capacity under chronic uncertainty. It replaces grit slogans with the RESILIENT nine-capacity model, concrete practices, and a weekly protocol covering identity boundaries, reappraisal, attention, decisions, recovery, social support, and antifragility.

Articles · 6 min read

RAG vs fine-tuning vs agents: three jobs, not a ladder

RAG, fine-tuning, and agents are three jobs. RAG, retrieval-augmented generation, looks up a named passage in a corpus you gave it, then writes. Fine-tuning changes model weights so a stable skill fires without fetching a document. An AI agent holds a goal, uses tools, and continues across steps under a written charter. They are not a ladder of sophistication. Fine-tuning does not make the model know Tuesday's PDF. RAG does not give the model a ceiling. An agent that cannot open a source still needs retrieval or a refuse. What is RAG owns the method. What is an AI agent owns the worker. This page is the comparison. The RAG Engineer is live. Fine-tuning win-rates are not this page.

Articles · 5 min read

Second-order thinking audit: the first-order metric is green

A second-order thinking audit is a written fail/pass on one decision before you commit. Four checks: who pays later, on which surface, whether you can reverse it this quarter, and which incentive the decision just created. A first-order win with an unnamed later bill is not a decision. It is a delay. CASCADE, on this site, is the personal habit that fires that pause in your own head. DOWNSTREAM is the organisational protocol for delayed, displaced, previously unowned costs. They are companions, not synonyms. What are downstream effects owns the definition. The operator article owns the longer map. This page is the card. Both books are in production and not for sale.

Articles · 6 min read

SENSE: a 90-day strategic intelligence layer

A strategic intelligence layer is a standing system that turns market, competitor, and internal signals into sourced claims a founder can act on. SENSE is one implementation in five moves: Scan, Evaluate, Navigate, Simulate, Execute. It has no Decide phase: the system prepares the choice, a person owns it. In ninety days you Diagnose, Deploy, then Evaluate-and-decide. Watchlist, assumption register, cadence, and kill rule are artifacts inside those gates, not a second calendar. The unit of work is a claim with a confidence grade, not a longer PDF.

Articles · 4 min read

STACK: version the agent OS like you version the app

STACK is a five-layer architecture for repositories where humans and agents share work: Structure, Toolchain, Agent configuration, Connection, Knowledge and quality. The agent operating system should be versioned like the application. If prompts, MCP servers, and skills live outside the repo, or live in five places inside it, you have agentic repo debt. The fix is ownership and a single source of rules, not a smarter model.

Articles · 5 min read

The Agentic Author: Stop Writing Books. Build a Production System.

The Agentic Author is the operator manual for BooksOS, a seven-stage nonfiction production pipeline where AI agents run stages and the author certifies quality gates. It covers evidence tracking, red-team review, bilingual native editions, and tool-agnostic design so authors ship serious work without surrendering the byline.

Articles · 3 min read

The Agentic Codebase: Version the Agent OS Like You Version the App

The Agentic Codebase is a production architecture book for technical founders and senior builders shipping agentic software. It introduces the STACK five-layer model so AGENTS.md, rules, skills, MCP, hooks, evals, and shell automation become versioned repository infrastructure that survives tool churn and team growth.

Articles · 8 min read

The Multi-Agent Organization: Close the Coordination Gap With CHORUS

The Multi-Agent Organization is an operator manual for 2-to-15-person teams that already mix human operators and specialized AI agent workers. It installs CHORUS, a six-skill coordination protocol: Charter, Handoff, Orchestrate, Review, Update, Sync. The book treats agents as workers with contracts, not as a chat demo or a vendor runtime, and it runs in the tools you already have.

Articles · 7 min read

Using AI agents effectively: one job, a ceiling, a log

An AI agent is already defined on this site: a system that holds a goal, uses tools, and continues across steps under a charter. Using one effectively means giving it one job, the smallest tool set that job needs, a written refuse-list, a hard ceiling on steps and spend, and a human on send, spend, delete, and speak-as-the-company. Writes should be draft-first or reversible, and an unfamiliar tool error should stop the run rather than start a retry loop. Optimisation is not a bigger model. It is a tighter charter, fewer write tools, a log you can open on Tuesday, full review of early writes, and a permanent human gate on irreversible calls. If it cannot refuse and cannot stop, it is not an agent you can run.

Articles · 4 min read

What are downstream effects?

A downstream effect is a consequence of an operational decision that is delayed, lands on a different surface than the metric you optimized, and had no owner in the room when the decision was made. Downstream consequences is the same river with slightly wider language. Downstream affect is the verb; the noun you want is effect. You do not need to predict every consequence. You need a way to name the next bend before commitment. The first-order metric can be genuinely green. Partner margin, a regulator queue, an agent network, or a customer renewal are typical landing surfaces.

Articles · 6 min read

What is a deal architect?

A deal architect is an operator who designs a high-stakes deal as a structure, not as a persuasion contest. The work is power, alignment, constraints, terms, and future options. Optimizing one visible number (price, close date, headline valuation) without those five is not architecture. It is a local maximum. The named system on this site is PACT. The book The Deal Architect is in production and not for sale. The definition is already usable.

Articles · 5 min read

What AEO means: answer engine optimization, not customs

AEO, answer engine optimization, is the craft of writing a page so an answer engine can quote a clean capsule: a direct answer near the top, named entities, visible sources, and no invented FAQ schema. GEO is the system around the crawl. AEO is the page. They fail together when the first screen is a brand manifesto. Authorized Economic Operator is a different AEO, in customs. Google ended FAQ rich results on 2026-05-07; a fake FAQPage is a liability, not a tactic. If the first screen cannot be quoted, the rest of the page will not save it.

Articles · 5 min read

What is an AI agent? A definition founders can use

An AI agent is a system that takes a goal, uses tools, and acts across more than one step without a human prompting each one. A chatbot answers a turn. An automation runs a fixed script. An agent is only accountable when it has a charter: inputs, outputs, an authority ceiling, and a review gate. Without those, you have a fluent wrapper, not an operator you can put on a roster.

Articles · 7 min read

What is an evidence ledger?

An evidence ledger is a claim-level register: what was asserted, which source supports it, how strong that source is, and what would retire the claim. It is not a bibliography dumped at the end, not a reading list, and not a folder of PDFs. The unit is one claim, not one source. The grade is a ceiling on the verb: an internal export can carry "in our data, this quarter" and cannot carry "in the market." The retire condition names something you would trip over in normal work, not "if it turns out to be wrong." Operators keep one so a number cannot wander between a chapter, a pitch deck, a back-cover line, and an answer-engine capsule. If the claim cannot point at a row, it is not ready to print, pitch, or paste.

Articles · 6 min read

What is an MCP connector?

An MCP connector is the attachment between a host (ChatGPT, Claude, Gemini, Grok, a desktop app, a CLI) and an MCP server. The server exposes tools, resources, and prompts. The host opens one client per server. Local connectors usually start a process over stdio. Remote connectors speak HTTPS (Streamable HTTP or SSE) so a cloud host can reach the server. ChatGPT custom MCP is remote-only. Claude Desktop can run local stdio servers from a config file. Grok custom connectors need a public HTTPS URL; localhost must be tunnelled. Gemini managed agents and Gemini CLI can attach remote MCP servers. Start first-connect diagnosis with transport, restart, and auth, not the model. Menus move. Recheck first-party docs. A connector without visibility, mutation, and a log is still a hole.

Articles · 4 min read

What calibration means: matching confidence to evidence

Calibration, on this site, is the practice of matching the strength of a judgment to the strength of the evidence, then updating when the facts move. Confidence is a report, not a personality trait. A loud claim with a thin source is a calibration failure, not conviction. The work is to score the claim, write what would change your mind, and revise the score when the evidence moves. It is not camera calibration, not model-weight calibration as a product claim, and not a psychometric test you can buy. The named system is CALIBRATE, in The Calibrated Mind, live in English and Dutch paperback and hardcover.

Articles · 6 min read

What is CHORUS? Coordination, not a choir

CHORUS is a six-skill coordination protocol for teams of 2 to 15 people who share work with specialised AI workers: Charter, Handoff, Orchestrate, Review, Update, Sync. A charter is seven fields. A handoff is a six-field packet. They are not the same artifact. It is not a choir, not an alphabet song, not CrewAI, and not a graph of model calls. If the receiver reconstructs the job from a chat log, you have a paste, not coordination. The book that owns the name is The Multi-Agent Organization, which is live.

Articles · 4 min read

What is RAG? When AI looks up a source before it answers

RAG, retrieval-augmented generation, is an AI method. A language model does not magically know your contracts, handbook, or last Tuesday's report. RAG makes it look up a specific passage in a collection you gave it, then write the answer from that passage. Retrieval is the product. The sentence is the last mile. If you cannot open the source behind the sentence, you do not have RAG. You have a chat window with extra text pasted in.

Articles · 5 min read

What is STACK? Five layers, not a vendor diagram

STACK is a five-layer architecture for repositories where humans and agents share work: Structure, Toolchain, Agent configuration, Connection, Knowledge and quality. Structure is the map an agent can navigate on minute one. Toolchain is the commands it may run. Agent configuration is versioned standing context. Connection is MCP with contracts and least privilege. Knowledge and quality is memory, evals, and CI, so a model upgrade does not silently lower the bar. It is not a vendor 'AI stack' diagram. The book that owns the name is The Agentic Codebase, which is live.

Articles · 7 min read

Why am I not mentioned in ChatGPT?

You are not mentioned in ChatGPT for one of a short list of operating failures, not because the model 'likes' someone else. There is no published ChatGPT citation ranking to buy. ChatGPT Search can only quote a public URL that OAI-SearchBot is allowed to fetch. GPTBot is training, not Search. ChatGPT-User is a person-triggered fetch. Testing a memory chat is not a Search test, and one run is not a result. Log three outcomes: absent, linked but unnamed, and apparently used without attribution. Absence points first to access and discovery. A source link without your name points to weak entity cues near the claim. Apparent reuse without attribution is not proof of copying; check whether the wording is distinctive and where the canonical version lives. How to get cited by ChatGPT is the longer operating note. This page is the diagnosis.

Articles · 6 min read

Why multi-agent systems fail: the coordination gap

Most multi-agent systems fail as coordination, not as models. A 2-to-15-person team adds specialised agents faster than it adds charters, handoff packets, review gates, and a shared view of state. Org charts name reporting lines. They do not name who holds context at the boundary. CHORUS is one protocol for that gap: Charter, Handoff, Orchestrate, Review, Update, Sync. A graph of model calls is not orchestration.

Projects

MindSesh

Browser-based, neurobiologically grounded sessions for focus, NSDR, breathing and stress recovery.

Projects

ReasonKit

A Rust-based reasoning engine for structured, auditable and verifiable AI workflows.

Projects

Undominated.ai

Independent AI inference price index. Capability first, then price. Names the models that are both worse and dearer. No cut of inference.

Concepts

AEO

AEO, answer engine optimization, is the craft of writing a page so an answer engine can quote a clean capsule: a direct answer near the top, named entities, visible sources, and no invented FAQ schema. GEO is the system. AEO is the page. They fail together when the first screen is a brand manifesto. Google ended FAQ rich results on 2026-05-07; a fake FAQPage is a liability, not a tactic. If the first screen cannot be quoted, the rest of the page will not save it.

Concepts

Agent orchestration

Agent orchestration is the assignment of work across humans and specialised agents so that handoffs, reviews, and state stay visible. A graph of model calls is not orchestration. Orchestration is who is chartered, what packet travels, what must be reviewed, and who sees the board. Without those four, adding agents adds waiting. CHORUS is one implementation of that operating layer. If you cannot point at the owner, the packet, the gate, and the board, you have a demo.

Concepts

AI agent

An AI agent is a system that takes a goal, uses tools, and acts across steps without a human prompting each one. A chatbot answers a turn. An automation runs a fixed script. An agent is accountable only when it has a charter: purpose, inputs, outputs, authority ceiling, success criteria, review, and escalation. Without those seven fields, it is a demo that can spend money. A wrapper around a chat model is not this. Neither is a scheduled script.

Concepts

BooksOS

BooksOS is a gated production system for serious nonfiction: Brief, Bible, Map, Draft, Edit, Package, Launch. Agents do specialist work inside the gates. Judgment, voice, and accountability stay human.

Concepts

CALIBRATE

CALIBRATE is the practice of matching the strength of a judgment to the strength of the evidence, then updating when the facts move. Confidence is a report, not a personality trait. A loud claim with a thin source is a calibration failure, not conviction.

Concepts

CASCADE

CASCADE is the personal cognitive companion to DOWNSTREAM: trace the next bend before you commit, in the meeting, in your own head. DOWNSTREAM is organizational. CASCADE is the habit that makes the protocol fire when nobody is watching. It is the pause that names who pays later, on a different surface, before you say yes. The book that owns this name is still in production; the habit is already usable. A first-order win with an unnamed later bill is not a decision. It is a delay.

Concepts

CHORUS

CHORUS is a six-skill coordination protocol for teams of 2 to 15 people who share work with specialised AI workers: Charter, Handoff, Orchestrate, Review, Update, Sync. Org charts name reporting lines. CHORUS names coordination lines.

Concepts

CITATION

CITATION is the standard for a page worth quoting: discoverable, attributed, answer-shaped, evidenced, and actually used. Search eligibility is not a citation. Ranking in ten blue links is a different job. Answer engines quote the first clean capsule they can verify. Invented FAQ schema is not a shortcut; Google ended FAQ rich results on 2026-05-07. The book that owns this name is still in production; the standard is already how this site is built. A mention without a checkable source is not this standard.

Concepts

DOWNSTREAM

DOWNSTREAM is an operational protocol for consequences that arrive later, on a different surface, with no owner in the room: Detect, Observe, Weigh, Navigate, Stream effects, Trace, Review, Effects on stakeholders, Assess systemic risk, Mitigate. It is an operating heuristic, not a validated risk instrument.

Concepts

Evidence ledger

An evidence ledger is a claim-level register: what was asserted, which source supports it, how strong that source is, and what would retire the claim. It is not a bibliography dumped at the end. Operators use it so a number cannot wander between chapters, books, or a back-cover line. If the claim cannot point at a row, it is not ready to print, pitch, or paste into an answer. The ledger is the product. The reference list is a rendering.

Concepts

GEO

GEO, generative engine optimization, is the work of becoming a source that answer engines can retrieve, interpret, extract, and verify. It is not a bag of hidden prompts. Eligibility, extractable answers, original value, and evidence that survives scrutiny are the levers. Ranking in ten blue links is a different job. Google does not require schema, chunking, or an llms.txt file for AI Overviews. A page that cannot be quoted cleanly will not be cited, even if it ranks.

Concepts

GRAIN

GRAIN is a retrieval loop: Gather, Rank, Assemble, Inspect, Navigate. Gather collects candidate passages. Rank orders them for the question actually asked. Assemble builds the context the model may see. Inspect checks that the cited chunk is the one used. Navigate decides the next lookup or the refusal. If you cannot point at the passage, you do not have retrieval. You have a fluent guess.

Concepts

Human in the loop

Human in the loop means a named person holds an authority ceiling and a review gate. It is not a human glancing at a dashboard. The human charters the work, can stop a tool call, and owns the outcome. If no one can fail the output against written criteria, the loop is decorative. A vendor slide that says HITL without a named reviewer is a slogan. The name, the ceiling, and the fail criteria have to live in the runbook, not in the deck.

Concepts

MCP

MCP, the Model Context Protocol, is a contract for how a model calls tools: what it may see, what it may change, and how the call is logged. The official split is host, client, and server, with tools, resources, and prompts as the primitives. Auth is optional; least privilege is not. A server without a contract is a hole in the repo. Founders should care about the interface and the permission, not the logo on the protocol. A clean server with no charter is still a hole with nicer plumbing.

Concepts

Override doctrine

Override doctrine is the written list of decisions an agent may never take: spending above a ceiling, speaking as the company, changing production data, or closing a judgment the founder still owns. It is the complement of a charter. The charter says what the agent does. The doctrine says what it must refuse. A refuse-list that lives in a Slack thread is not doctrine. It is a hope. Write it before the first tool that can send, spend, or delete.

Concepts

RAG

RAG, retrieval-augmented generation, means the model must fetch named passages before it answers. The retrieval is the product. The generation is the last mile. If you cannot inspect which chunk was used, you do not have RAG. You have a language model with extra context stuffed in the prompt. A vector store is a component. GRAIN is the loop that keeps the answer honest. Adding embeddings is not a finished system. Naming the passage is.

Concepts

RESILIENT

RESILIENT is nine trainable capacities for founder pressure: reappraisal, emotional granularity, somatic regulation, identity flexibility, learning orientation, interpersonal support, executive function, narrative integration, and titrated exposure. Resilience is infrastructure, not a personality trait.

Concepts

ROSTER

ROSTER is a six-skill protocol for a one-person company that runs named specialist agents: Roles, Objectives, Skills and tools, Triggers, Evaluation, Rotation. Roles give each agent a charter. Objectives make the charter measurable. Skills and tools bound what the agent may use. Triggers decide when it runs. Evaluation scores output instead of vibe. Rotation retires the agent that was built for last quarter's work. One chatbot that never hands off is not a team.

Concepts

ROUTE

ROUTE is a five-skill discipline for governing a mix of language models: Register, Objective typing, Utilize policy, Track, Evolve. Register catalogs every model with mandate, cost class, privacy class, and retirement status. Objective typing names what a task actually requires before a model is chosen. Utilize policy writes the routing rules, cascades, and fallbacks. Track measures cost, latency, and quality per route. Evolve promotes, demotes, and retires on a cadence. A leaderboard pick is a collection. ROUTE is a portfolio.

Concepts

SENSE

SENSE is a five-move strategic intelligence layer: Scan, Evaluate, Navigate, Simulate, Execute. It turns agent work from a productivity shortcut into a sourced claim that changes a watchlist, an assumption register, or a Monday decision.

Concepts

STACK

STACK is a five-layer architecture for repositories where humans and agents share work: Structure, Toolchain, Agent configuration, Connection, Knowledge and quality. The agent OS is versioned like the app.

Concepts

Strategic intelligence layer

A strategic intelligence layer is the standing system that turns market, competitor, and internal signals into sourced claims a founder can act on. It is not a dashboard wall and not a weekly PDF. The test is whether a claim changes a watchlist, an assumption, or a Monday decision with a confidence grade attached. SENSE is one implementation: scan, evaluate, navigate, simulate, execute. If competitive work is still a founder reading Slack at midnight, you do not have a layer. You have a habit that will miss a week.

Concepts

Agentic Codebase

Repository architecture that keeps AI-agent work production-safe: layers, review gates, and durable interfaces that survive real deployments.

Concepts

AI Agents

Architectures and operating patterns for agentic AI systems that founders and engineers can deploy without cargo-cult complexity.

Concepts

Downstream Effects

Delayed, displaced consequences of operational decisions: second-order effects, partner and regulatory landing surfaces, and the audit that makes the next bend visible.

Concepts

Founder Resilience

Operator-grade resilience systems for high-stakes founders who need clear thinking under pressure, not motivational slogans.

Concepts

Marketing

Choice architecture, persuasion literacy, and honest commercial practice.

Concepts

Multi-agent organization

Small teams where human operators and specialized AI workers share goals, handoffs, and review gates.

Concepts

Positioning

Positioning as a decision about who you refuse and what you will not claim.

Concepts

TRACE

TRACE is a five-step research loop: Target, Retrieve, Assess, Cross-check, Export. Target locks stakes, scope, and output format before a search runs. Retrieve gathers independent sources instead of one confident pass. Assess tiers each source so weight follows credibility. Cross-check requires a second source before a load-bearing claim can enter the memo. Export is a claim ledger a co-founder can act on without redoing the work. Fluent collection is not research. TRACE is the verification structure.

Books

The AI Risk Register

Build a Living Register That Names, Owns, and Tests Every AI Risk Before It Names You

Books

The AI Team Dynamics

Designing Mixed Human and Agent Teams So Productivity Compounds and Accountability Stays Human

Books

The Crisis Command

COMMAND: Calm, Assess, Map, Mandate, Unify, Notify, Decide. The Trained Response

Books

The Judgment Gym

Maintain the Thinking That Makes You Worth Hiring When AI Thinks Faster

Books

The Knowledge Scale

Engineer Organizational Memory So Your Company Stops Forgetting What It Already Learned

Books

The Meaning Operator

Applied Existentialism for Founders: Why You Build, What It Costs, What It Is Worth

Books

The Pricing Mind

Behavioral Economics of Price: How Buyers Experience Value and How to Design That Experience

The knowledge atlasLEN P. VAN DER HOF

Use the index for keyboard navigation. On the map, use arrow keys to pan, plus and minus to zoom, and Escape to close details.

Export map
JSON

Featured research

Start with the strongest signal.

Latest

Recent articles.

Content pillars

Explore by operating problem.

Portrait of Len P. van der Hof

Written by

Len P. van der Hof

Len P. van der Hof builds practical AI systems, digital ventures and evidence-informed tools for founders.

Author profile
Markdown for LLMs