A Polished Document No Longer Proves Anyone Thought. Test the Author Instead.
Watch the video version:
A principal product manager presented a six-page document at an Amazon review, and by one attendee's account, everyone in the room could see he was "reading the document for the first time." Steve Huynh, who spent almost twenty years at Amazon, the company famous for running decisions on written narratives instead of slides, passes the story along in a recent essay on the decline of that writing culture. It came to him from someone in his community who sat in that room.
That is the problem Huynh is pointing at. "With AI, you can trivially produce six pages, no critical thinking required," he writes, and he describes a senior engineer who has started tagging documents "AI;DR."
Writing six clear pages used to take days, and nobody could do it without working through the problem, so a polished document was decent evidence that someone had thought. Teams learned to read polish as thought, and for years that was a sound read. That signal is gone. No amount of careful reading now tells you who thought and who prompted.
Ask engineering leaders what to do about it and you hear one of two answers. Keep humans writing the documents that matter, which is close to Huynh's own fix: human authorship for the decisions that are hard to reverse, with the author owning every word. Or stop worrying: a good document is a good document, whoever or whatever wrote it. Both keep the review on the page. The first tries to make the page trustworthy again, and the second declares the page enough. Neither asks whether the person proposing the decision can defend it.
When the page stops telling you anything, the decision falls to the most senior reader. That person's skim, gut, and one question decide, and the proposal passes on how that reader reacted, not on what the author understands. Authors notice. They write for that reader and add polish, because polish keeps passing. Every proposal waits on one person's attention, and that person becomes the constraint on how fast the team can decide.
The larger cost arrives later. A decision approved on the page ships with no owner who knows where it is weak. Nobody asked what breaks at ten times the load or what would make it wrong, so nobody is watching for it. When conditions change, the decision expires quietly and stays in force, and the reversal comes late, when it costs more.
Neither a stricter AI policy nor a closer read fixes this. The review has to move from the page to the person.
I call it the Author Test. I ask open-ended questions about any proposal, and I focus on four: How does this scale? What flaws do you see? Where are the stress points? What change in conditions would make you revisit this decision? No document answers them for its author. Someone who did the thinking can answer them slowly, in their own words, and still answer well. Someone who only prompted the pages runs out of answers at the first follow-up. Because the questions are written down, anyone on the team can ask them, and the review stops waiting on one senior reader's skim.
Three rules keep it from turning into a performance. Send the questions with the request, so the author knows them before writing a word. Preparation then beats quick talking, and the quiet engineer who did the work arrives with answers. Written answers do not pass on their own, since a model can draft those too. In the review, ask one follow-up under each answer, such as why it breaks at that point or what that flaw would cost, and the author has to answer it without the page. Scale the test to the cost of undoing the decision: one that is expensive to reverse gets all four questions, and a cheap one gets only the fourth. Record the fourth answer with the decision, as a condition someone can watch for, such as a volume, a date, or a vendor term, and name the person who watches for it, so someone notices when the decision has expired. A proposal whose author cannot answer goes back, however good the pages look.
Some answers to the stress-points question will not resolve. Two legitimate sides each need something the design cannot give both, both are right, and no rewrite of the proposal makes that go away. That is a Structural Bearing™. Put it in the record by name, where the next decision can find it, instead of smoothing it over.
Start with your own proposals. If the leader's ideas skip the test, it is a rank ritual, and the team will see it for one. Every time a polished document passed because it looked right to you, you were the seniority filling the gap. The broken signal put you there; the four questions take you out.
Take the last proposal you pushed through yourself. Without opening the document, answer the four questions about it. Then do the same for the last proposal you approved from someone else. If you cannot answer the fourth for either one, nobody is watching that decision, and the page is what approved it. Send the four questions with the next proposal request, and answer them first for your own.
The Edge Case walks through the Structural Bearing™, the tension a stress-points answer sometimes uncovers, and how to state it in two sentences, what both sides require and what makes the gap permanent: https://TheEdgeCaseBook.com
Years of argument over a tension that will not resolve, compressed into two sentences: what both sides require, and what makes the gap permanent.
