You become the ceiling. Here is the exact question this article answers: why does undocumented judgment quietly cap a business, and what changes when you finally write it down?
The Question I Could Not Answer
A few years ago somebody on my team asked me a simple question about how we decide which prospects to spend real time on.
I opened my mouth to answer and nothing usable came out. I said something about feel. I said something about fit. I said the phrase “you kind of just know,” which is the single most useless sentence a leader can say to a person trying to learn.
I had been making that decision well for years. I could not explain it. And in that moment I understood that every good decision in my company was sitting in exactly one place, which was my head, which meant every good decision in my company had to wait for me.
Here is the direct answer to the question in the headline. When everything important lives in your head, you stop being the founder and become the bottleneck. The business cannot grow faster than you can personally attend to it, and, more painfully, nobody else can get good at the thing you are good at, because you never made it learnable.
That was true before AI. It is more expensive now, because the tools that would multiply that judgment cannot see it either.
Key Takeaways
- Undocumented judgment makes the founder the ceiling, because every good decision has to route through one person.
- Model access has stopped being a competitive advantage now that open-weight models like Kimi K3 and Inkling-Small match or approach frontier performance.
- The differentiator moved upstream from which model you use to what context you can feed it.
- Anthropic’s engineering research treats context, not intelligence, as the actual bottleneck in AI systems.
- Writing down how you decide is uncomfortable because the parts you do best are the parts you cannot immediately explain, and that is exactly why it is valuable.
The Part Nobody Warns You About
The reason nobody documents their judgment is not laziness. It is that expertise hides from the expert.
The things you are best at have become invisible to you. You do not experience them as decisions anymore. You experience them as noticing. Somebody describes a prospect and you already know. Somebody shows you a draft and you already know what is wrong with it. There is no visible reasoning step because the reasoning compressed years ago into something that feels like instinct.
So when you sit down to write it out, you get two sentences and then a wall.
I hit that wall repeatedly. I would start a document called something confident like “How We Qualify Leads,” write four bullet points that any competent person could have guessed, and then stall on the part that actually mattered. The four bullets were not my judgment. They were the generic version of my judgment. The real thing was in the exceptions, and I could not surface the exceptions by staring at a blank page.
What eventually broke it was giving up on writing and switching to talking. I had somebody interview me about specific past decisions, one at a time, and push every time I said something vague. Why that client and not the other one. What did you notice. What would have changed your mind. Twenty minutes of that produced more usable material than months of good intentions.
I am telling you this because I want you to expect the wall. If you sit down to document your judgment and immediately feel like you have nothing to say, you are not unusual and you are not shallow. You are just experiencing what it feels like when knowledge is compressed. Talk it out instead. Have somebody ask you about real cases. The rules are in there.
What Changed This Summer
Here is why this stopped being a nice-to-have and became the whole game.
The largest open model in history shipped, and it is free. Moonshot AI released Kimi K3 in mid-July 2026, a 2.8 trillion parameter model, and completed open-sourcing it by the end of the month. Kimi K2.6 already benchmarks alongside Alibaba’s Qwen 3.6 Max and DeepSeek V4, sitting just behind top closed models. Your competitor can download that today.
Small open models are matching big ones. Thinking Machines Lab released Inkling-Small with 12 billion active parameters. It matches its full-size sibling and beats it outright on reasoning and agentic coding tests. Twelve billion active parameters, running on hardware a small business can actually reach.
Frontier capability is arriving under permissive licenses. DeepSeek V4 shipped with an MIT license and a one million token context window. There is no gatekeeper left to be on the right side of.
The bottleneck moved, and Anthropic said so directly. Anthropic’s engineering team has published extensively on context engineering, framing context as the bottleneck rather than intelligence, and describing the core discipline as optimizing signal-to-noise ratio within a token budget. Their own research across roughly 400,000 Claude Code sessions found that domain expertise amplifies how much useful work the model produces per instruction. The person who knows the business well gets far more out of the identical model.
Put those together and the conclusion lands somewhere uncomfortable and clarifying. Two businesses can run the same model on the same prompt and get wildly different results, and the difference is not talent. It is what the model was allowed to know.
So the thing sitting undocumented in your head is not just a succession risk anymore. It is the input that would separate your AI output from your competitor’s, and it is unavailable.
What I Actually Did
I stopped trying to document everything and picked one decision.
I chose the one where people most often told me we were better than alternatives, because that is where the real judgment was hiding. Then I had someone interview me about five specific past instances, including two where I got it wrong.
The wrong ones did more work than the right ones. When you explain a good decision you tend to narrate the rule. When you explain a bad decision you have to explain what you missed, and what you missed is where the exceptions live.
Out of that came a document. Not elegant. Roughly a page and a half of if-then rules, a list of named exceptions, and a section for what to do when none of the rules apply. Then I ran it backward against five older cases where I already knew the outcome, and it got two of them wrong. So I fixed those two specific things and nothing else, because the temptation to keep polishing is real and it is how documents die.
Then I loaded it as standing context for the AI tools we use.
The change in output was not subtle. Same model. Same requests. Suddenly the recommendations sounded like they came from somebody who worked here.
That was the moment the whole thing clicked for me. I had spent a year trying to get better output by getting better at asking. The actual answer was to stop being mysterious.
Practical Steps
1. Pick one decision, the one you are known for. Not the one that annoys you most. The one where customers say you are better. That is where your undocumented advantage lives.
2. Do not write. Talk. Have somebody interview you, or use AI as the interviewer, about specific past instances of that decision. Insist on real cases, not hypotheticals.
3. Include the ones you got wrong. Two failures will teach you more about your own rules than ten successes. Failures force you to name what you overlooked.
4. Convert what comes out into if-then rules plus named exceptions. Then add a section for situations none of the rules cover, and say honestly what you do then. That section is not a weakness. It is the most honest part of the document.
5. Test it backward against five known outcomes. Run past cases through the rules and see where the rules would have been wrong. Fix only what actually broke. Resist improving things that were not tested.
6. Load it as standing context, not as a one-time prompt. A document you paste occasionally is a habit that will fail. A document your AI reads every time is infrastructure.
7. Put a quarterly review on the calendar. Businesses change. A context file nobody maintains becomes confidently wrong, which is worse than being absent.
Frequently Asked Questions
How long does documenting one decision actually take?
The interview takes about thirty minutes. Converting it into rules takes about an hour. Testing it against past cases takes another hour. Plan for half a day and you will be close, which is dramatically less than the years most founders spend intending to do it.
What if my judgment really is just intuition?
Intuition is compressed pattern recognition, not magic. It can be decompressed by examining specific cases rather than trying to state general principles. If you can explain why you decided a particular case, the rule is in there.
Does this mean my competitive advantage disappears once it is written down?
The opposite. Unwritten judgment can only be applied by you, which caps it at your available hours. Written judgment can be applied by your team and your tools. Your competitor still cannot see it, because it lives in your systems, not on your website.
Which decision should I document first if several feel equally important?
The one that happens most often. Frequency compounds. A moderately valuable decision made weekly returns more than a highly valuable decision made twice a year.
How is this different from writing a standard operating procedure?
An SOP documents steps. This documents choices. Most SOPs describe what to do and skip entirely over how to decide which thing to do, which is where all the actual value sits.
The Close
“You kind of just know.”
I still hear myself saying it sometimes, and it still makes me wince, because I understand now what I was really telling that person. I was telling them they could not learn this. I was telling them to come find me every time. I was, without meaning to, making myself the ceiling of my own company.
There is a version of founder identity that quietly enjoys being indispensable. I am not going to pretend I was above it. It feels good to be the one who knows. But indispensable and irreplaceable are not the same word, and only one of them is a compliment.
Your competitor can download a model within a few points of yours by dinner tonight. They can hire your playbook and copy your website. What they cannot download is how you actually decide, and right now, neither can your team, and neither can your tools.
So pick one decision this week. The one you are known for. Sit down with somebody and talk it out until the rules fall out of your mouth.
You have been carrying something valuable in a place nobody else can reach. Put it somewhere it can work without you.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs turn AI from a novelty into an operating system for their business. He is the creator of the Perfect Prompt Framework, a speaker, and a father of eleven, which turns out to be excellent training for realizing you cannot personally be present for every decision.