The most valuable file at Flip Education is a rules document. About 18,000 words of it now. Product ships from other files. This one carries the judgment.
Here's the kind of thing it exists to prevent. In May, a sync monitor died quietly behind an SSL error and stayed dead for seven days while its dashboard looked fine. Same month, a nurture email went out in English to a teacher in Chile because one helper function got skipped. Neither bug was hard. What was hard was noticing, because nothing anywhere was obligated to notice. So the file now says: every pipeline names the query that proves it's alive, and every plan says what fires when it breaks. Boring rules. They work.
Give a capable agent a task and you'll get a competent answer fast, usually the first plausible one, and that's how you lose a week without noticing. A senior operator earns their pay in the discipline they apply before committing to an answer. Write that discipline down and every agent in the company inherits it the same afternoon.
Here's some of what mine says. Not the file, the spirit of it.
Nothing non-trivial gets planned from a blank page. The file demands ten minutes first on how the best teams already handled this class of problem, and someone at Stripe or Linear has usually written it up. Then comes the translation step, which matters more than the search itself: keep the one load-bearing idea and drop everything that assumes a 400-person company. I'll admit most of those ten-minute searches confirm what a decent operator would've guessed anyway. The tenth one saves a month. No way to know in advance.
In May I pointed four research agents at a market question, framed the way I already believed it. They came back agreeing with me, with sources. Which proved nothing. Framed the other way, the same agents would've built an equally convincing pile for the opposite answer, because search rewards whoever's asking. The rule that came out of that week says any strategic question starts with three or four explanations that can't all be true, plus a written note of what finding would kill each one. Intelligence analysts have worked this way since Richards Heuer wrote it up for the CIA in 1999. It triples the research bill on questions like this.
The May failures shared one root cause. Nothing was obligated to notice them. So reliability questions now sit in every plan on the same line as the feature itself: what breaks loudly when this fails, and which query proves it ran today. Tests can't answer that. Tests catch the failures you imagined at your desk, and both of May's were failures nobody imagined. Standing agents also get a watcher on their own output, tuned to raise a hand when the numbers stop moving for a day.
When an agent says done, it's describing what it meant to do. In May a survey of what had shipped came back confident, and I spot-checked three of its four claims. All three held. The fourth, the one I skipped, sat on six commits and 865 lines I'd never looked at. Sampling felt rigorous and it was closer to luck with a clipboard. So the rule is blunt now. Read the diff before believing it, and open the artifact the way a user would.
Some work is a dial. Spacing on a page, the tone of a letter. For that kind, the rule is one change, then one look at the rendered result, every time. A spacing tweak gets a fresh screenshot at phone width; for tone, I re-read the whole letter as the PDF a reader would open. It turns a five-minute edit into a twenty-minute session, which is annoying, and I keep it anyway. Every polish disaster I've been asked to review came from somebody turning three knobs between looks.
Prose rules decay. Six months on, a fresh agent stops following the paragraph nobody enforces, and nothing tells you it stopped. So any lesson worth keeping gets promoted out of prose into a mechanism, something like a validator that rejects the file or a block on the send button. My outbound tooling won't release another batch once bounces cross 3% in a 50-send window. Nobody remembers that rule anymore, and nobody has to. That's the point of building it. Each gate costs an afternoon. The lesson it encodes usually cost a week, so the arithmetic isn't close.
None of this is exotic. Most of it is what a strong operator does without thinking, written down precisely enough that a machine can do it too, and that writing turns out to be the real job of running a company on agents. The models arrive capable, and the file is what makes them careful.
This file is the discipline layer behind the practice. The gates that enforce it are on GitHub.