Anyone who has run a team has had this moment. A new hire asks why something is not done a certain way, you open your mouth to explain, and two sentences in you realize you cannot articulate it. So you say: we just don't do it that way here.
That sentence carries your team's entire body of judgment. It has never been on any balance sheet and it has never been written down.
That used to be fine. People taught people, and in three to six months it transferred. It is a problem now, because anything you have not written down is something AI cannot learn.
Anthropic published a startup guide last week from interviews with 15+ fast-growing companies. Plenty of people have summarized the five rules. What stopped me was elsewhere.
Victor Hunt, co-founder of Zingage, said they gave Claude full autonomy early on, then started writing down everything that could not move: how they frame problems, what has to be true no matter what. It came to 567 lines describing how that team thinks.
I went and counted my own file. 580 lines.
I have never met Hunt and we work on different things, yet we landed within 13 lines of each other. A team's non-negotiable judgment seems to have a natural size, somewhere around five or six hundred lines. Write less and it is slogans. Write more and nobody reads it.
Those 567 lines are an asset. It never appears in financials and you cannot put it in a job description, but it determines how fast a new hire becomes useful, how often work gets redone, and whether decisions hold when the founder is not in the room. That asset used to live only in senior people's heads, and when they left it was written off.
The part that actually matters is what this implies about bad teams.
If your standards are unclear, AI does not fix that. It executes the unclear standard faster, in more places, at lower cost. Every ambiguity you never resolved now gets reproduced at machine speed.
Artemis Security ships 6,000+ PRs a week. The number reads like agents made everything fast. Co-founder Dan Shiebler puts the causality the other way: that deployment speed only works because they invested deeply in testing infrastructure. Speed was not granted by the agents. Verification came first, and only then was speed safe to release.
That is why, of the five rules, the one I would keep is Trust but Verify. Without reliable verification, automation does not raise your speed. It raises the speed at which you are wrong.
Our own version was small and embarrassing. We had a check for AI-sounding phrasing before publishing here, and it printed a warning. Drafts with five em dashes went out anyway, because a warning depends on a human noticing. As a hard gate it changed nature. A gate depends on nobody.
The uncomfortable read of this entire guide: AI is a multiplier on whatever your organization already is. Good judgment gets amplified. So does the absence of it.
Top comments (0)