Resources
The sources behind how we work
Our claims should not rest on our word alone. These are the people we learned from, each with a line in our own words, and a link where one exists.
Same model, different answers
Every AI setup has three parts: the app, the model and the context. Here, the context changes the answer.
The three parts →-
Tim O'Reilly, founder of O'Reilly Media · Interview on AAIF Live · 2026
Asked if an overloaded team should hire, the AI said yes. Pointed at different expertise, the same question said restructure the team instead.
-
Julie Zhuo, Sundial.ai · You Cannot Mandate an AI Transformation · The Looking Glass · 2026-09-01
By her own figures for her own product, the same questions scored just over 80 percent, then 98 percent once the missing context was added.
-
Sean Goedecke · LLMs Reward Expertise · seangoedecke.com · 2026-07-24
Everyone talks to the same model, but Goedecke argues that knowing the field is the most important skill in prompting.
-
Zach Davis, Randy Shoup, James Wickett and others · Agentic AI and Code Reviews · IT Revolution, Enterprise Technology Leadership Journal · 2026
The authors argue that agents need precedent, such as why a similar approach was rejected before, and not only rules.
-
Ash Maurya · Building High-Converting Landing Pages with AI, Part 1 (Strategy) · LEANSpark live session · 2026-05-28
Maurya used the same prompt each time, and the page scored far higher once he added his brand, interviews and testimonials.
What was never written down
Most of what an organisation knows was never written down. We call that dark context.
What is dark context →-
K. Sudhir, Yale School of Management · Design Agentic Systems Around the Implicit Rules That Govern Your Company · Harvard Business Review · 2026-06-19
Sudhir argues that you can set up an agent only with a company's written rules, not with its unwritten judgment.
His questions for any role: What do people notice that is not in the data? What do they care about beyond the job description? When do they slow down?
-
Dearstyne and others, Microsoft Research and University of Notre Dame · Revealing the Dark Matter · ICSE 2026, New Ideas and Emerging Results · 2026
Building an agent for software testing, the team found that some of the engineers' expertise was never put into words at all.
-
Thinking Machines, shared by Mira Murati · The Future Worth Building Is Human · Thinking Machines blog · 2026-07-10
The essay argues that useful knowledge is tacit, local and held by the people who gained it, so one central model will struggle to gather it.
-
Qadri, Denton and others, Google Research and Google DeepMind · Working with Agentic Teammates · arXiv · 2026-09-24
In this interview study, people on 11 teams said one agent misread their unwritten rules, so it disrupted their work and wasted their time.
-
Robin Teigland · Organizational Blindsight · LinkedIn · 2026-10-18
Teigland argues that work flows through informal networks that leaders do not see, so they will also miss how AI changes them.
Why people share, and why a room
People share what they know when they choose to, not when they are told to. That is why we work in a room.
Why a room →-
Margit Osterloh and Bruno S. Frey, University of Zurich · Motivation, Knowledge Transfer, and Organizational Forms · Organization Science · 2000
Tacit knowledge moves when people want to share it, not by contract, command or price. In their example, a firm wrote its knowledge down for outside suppliers, and its rivals could then copy it.
-
Matt Beane · The Skill Code · HarperCollins · 2024
Skill passes from expert to novice in shared work, and Beane shows how intelligent tools lift the novice out of that work.
-
Donald A. Schön · The Reflective Practitioner · Basic Books · 1983
Schön shows that professionals know more than they can say, and become aware of it by putting it into words for others.
Keeping it current
Written down does not mean up to date. That is why our review comes back every month.
The review →-
Karri Saarinen, CEO of Linear · Retire the Word Context · X · 2026-11-04
Saarinen argues that context is knowing which parts matter, and that an old, clear explanation can mislead more than a missing one.
-
Joao Wenzel, COO of Citadel Securities · Your Company Has No Model of Itself · talk at AI Engineer NYC · 2026-10-14
Wenzel's team keeps a monthly record of its projects in which every claim carries a date and ages.
-
Tobi Lütke, CEO of Shopify · Learning on the Shop Floor · X · 2026-05-09
People watched Shopify's coding agent work in public channels and wrote down what it should have known, and it got better.
-
Stewart Brand · on The Ezra Klein Show, about his book Maintenance: Of Everything, Part One · 2026-04-24
Brand argues that you truly own only what you know how to keep working.
Whose knowledge is it?
Who owns what a company knows? Our answer: you decide what goes in and what stays yours.
What stays dark →-
Tim O'Reilly · the same interview, as above
O'Reilly argues that human expertise trained the models, and that the AI economy still needs rules for who gets paid for it.
-
Satya Nadella, chairman and CEO of Microsoft · The Reverse Information Paradox · X · 2026-07-12
Nadella argues that a company pays for AI twice: with money, and with the knowledge it must reveal to make the AI useful.
-
Raffi Krikorian, CTO of Mozilla · Can AI Keep a Secret? · Owners, Not Renters · 2026-09-02
Krikorian reads early US court rulings to mean that a business secret built inside a consumer chatbot may lose its protection.
Why one person gets ahead and the team does not
Often one person gets far with AI while the rest of the team waits.
One person has it working →-
Boris Cherny, Claude Code at Anthropic · Steps of AI Adoption · X · 2026-07-16
Cherny's steps of AI adoption put the limit at each step on people and the organisation, not on the model.
-
Benedict Evans · The Widening AI Perception Gap · Threads · 2026-07-23
Evans argues that the gap in AI use between tech and everyone else is widening, because most people and companies do not quickly rebuild how they work.
-
Benedict Evans · AI, Tools and Transformation · ben-evans.com · 2026-09-03
Evans argues that the hard part is not making a tool, but knowing that you need one and what it should do.
-
Cassie Kozyrkov · The AI ROI Metric Trap · LinkedIn · 2026-05-19
Kozyrkov argues that AI goes to work that is easy to measure, and companies let go of the people who hold unwritten context.
Where we might be wrong
These cut against parts of our view, and we keep them because an untested view is worth little.
-
Vas, CEO of Varick · AI Adoption Is a Myth · X · 2026-08-07
Vas argues that at every client a few people become heavy AI users and most do not, and no rollout changes that split.
We take from it that our work must not depend on every person becoming a heavy user.
-
Bridgewater AIA Labs with Thinking Machines Lab · Learning to Replicate Expert Judgment in Financial Tasks · Thinking Machines · 2026-06-30
Prompts written by experts raised accuracy to the mid-70s, then stalled below the 80 percent the investors need, because the judgments that count most are often the hardest to explain. Their answer was to train a model on the experts' own labels.
We take from it that words have limits, and a room will not put everything into words.
-
Lin Qiao, CEO of Fireworks AI · Owning vs. Renting Intelligence · X · 2026-06-15
Qiao argues that a company should own a model tuned on its own work, because control matters more than cost. It is a vendor's argument: her company sells training and hosting for open models.
We take from it that the context may not be the only part a company can own.
Does one of these sound like your team?
Book a call and tell us where you are.